GEO: how to get cited by AI answers

This is guide 5 of 5 in our SEO learning path, and you can read it straight after Guide 2. It takes one question, "can you get us into ChatGPT?", and answers it with what the search engines have said, what the studies measured, and what you can check on your own site.

It's written for in-house SEOs and content leads, agency staff who get asked that question, business owners deciding whether to pay for "GEO services", and developers who look after robots.txt. There are nine chapters, about 80 minutes in all. Read chapters 1 to 3 first for the evidence, then work through chapters 4 to 8 with your own site open, and keep chapter 9's checklist and 30-day plan for last.

Every number comes with its source, date and sample, and vendor studies are marked as vendor studies. The stories from our own work (a sports platform, a perfume brand, and the students I taught) say only what happened, and none of them claims an AI citation.

A client asks, "Can you get us into ChatGPT?" Before you answer, you need two things straight: what the words mean, and what the search engines themselves have said. This chapter covers both. The rest of the guide is the practical part.

Where the words come from

GEO (generative engine optimization) is the work of making your content more likely to be used and cited when an AI system writes an answer. The term comes from a research paper, "GEO: Generative Engine Optimization", by Pranjal Aggarwal and colleagues from IIT Delhi, Princeton University and two independent researchers. It went up on arXiv in November 2023 and was accepted at KDD 2024.^[1] The first author was at IIT Delhi, so the name of a whole industry trend started partly in India.

The other labels mean roughly the same job. AEO (answer engine optimization) puts the focus on "answer engines" such as ChatGPT and Perplexity. LLMO (large language model optimization) is a third variant. Patrick Stox of Ahrefs titled a 2025 conference talk "GEO? AEO? LLMO? What's With All This AI Stuff?", which tells you how settled the vocabulary is.^[2] Google's own guidance on third-party services mentions both AEO and GEO by name.^[3]

We use "GEO" in this guide because it has a paper behind it. Don't let anyone sell you AEO and GEO as two separate services.

What changed for searchers

For most of search's history, you typed a query and got ten links. Between 2024 and 2026 that changed fast:

• May 2024: Google's AI Overviews rolled out to all US users.^[4]
• 31 October 2024: OpenAI launched ChatGPT search; everyone could use it from 5 February 2025.^[5]
• 24 June 2025: Google's AI Mode came to India in Search Labs, the first market outside the US, with the Labs requirement dropped weeks later.^[6]
• 20 May 2026: at Google I/O, Google brought AI Overviews and AI Mode together into one AI Search experience, live that day worldwide.^[7]

The scale is large. Sundar Pichai said in July 2026 that AI Mode had passed 1 billion monthly active users and the Gemini app had 950 million.^[8] OpenAI said in February 2026 that ChatGPT had reached 900 million weekly users.^[9]

What it means for clicks

Two facts are true at the same time, and you need both when you talk to a client.

First, AI answers take clicks. In Pew Research Center's study, people clicked a link inside the AI summary itself in just 1% of visits. That's US browsing data, not a global figure. Second, AI assistants are huge products but small referrers. The SE Ranking figure is from a vendor, and its sample is sites connected to Google Analytics. So GEO is mostly about being seen and named, and only partly about traffic. Chapter 8 covers measuring it.

Google's position, in its own words

On 15 May 2026 Google published a guide called "Optimizing your website for generative AI features on Google Search", announced by John Mueller.^[12] It is the clearest statement any search engine has made on GEO. The key line:

From Google Search's perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.^[13]

The guide says Google's AI features are "rooted in our core Search ranking and quality systems". It then lists things you can ignore for Google Search:^[13]

• llms.txt files and other "special" markup or Markdown files (chapter 7)
• "Chunking" your content into small pieces
• Rewriting content just for AI systems
• Seeking inauthentic "mentions"
• Over-focusing on structured data

What it asks for instead is "non-commodity" content: things only you can write, from first-hand experience. Its example contrasts a generic "7 Tips for First-Time Homebuyers" with a homeowner's story about inspecting a sewer line. Chapter 4 shows how to apply that.

The older "AI features and your website" page is even shorter. It says there are no additional requirements to appear in AI Overviews or AI Mode. A page has to be indexed and eligible to show with a snippet, and there's "no special schema.org structured data that you need to add."^[14]

The rules that changed in 2026

Two changes make the "anything goes" GEO tactics risky.

On the same day as the new guide, Google clarified that its spam policies cover AI answers too. Spam now includes "attempting to manipulate generative AI responses in Google Search".^[15] As we read it, hidden text written for bots, fake mentions and mass pages built to catch AI answers now fall under the same policies as old-school spam.

Then on 5 June 2026 Google published guidance on third-party SEO tools and services, naming "AEO" and "GEO" tools among the things to look at carefully. Its words: "Third-party tools don't have access to our internal ranking data." And: "They can't guarantee performance."^[3] The AI optimisation guide says the same about AI systems.^[13]

The Search Console setting you should know about

Google's guide adds one condition: to be eligible for its generative AI features, a site must be "included in Search generative AI features in Search Console".^[13] That's a toggle Google announced in June 2026. It decides whether your site appears in AI Overviews, AI Mode and AI Overviews in Discover, and helps ground their answers. It rolled out to all sites worldwide by 31 August 2026.^[16] Every site is included by default.^[17] Chapter 6 explains what switching it off does. If you're cited nowhere in AI Overviews, check this setting first. It takes two minutes.

Bing's position

Microsoft goes further than Google in naming the term. In February 2026 Bing rewrote its Webmaster Guidelines^[18] to cover Copilot answers as well as normal results. As Search Engine Journal reported it, the guidelines now name "Generative Engine Optimization (GEO)". They treat it as making content eligible to ground and be cited in AI answers, and say GEO doesn't guarantee citations, just as SEO doesn't guarantee rankings. They also turned a one-line warning about prompt injection into a full section on prompt injection and AI manipulation.^[19] In May 2026 Bing's search team wrote that the index is moving "from ranking pages to supporting answers". They describe the unit of value as "groundable information": specific facts that can be supported and traced to a source.^[20]

Keep that idea in mind for the whole guide. Clear facts with a clear source are what an AI system can use.

Where we stand

GEO is SEO plus brand building plus being cited on other sites, measured with some new reports. Lily Ray, who tracks AI citations for her own studies, put it plainly in January 2026:

AEO/GEO is not an overhaul or abandonment of SEO.^[21]

Patrick Stox's slide said it in five words: "AI Search Is Still Search."^[2]

Our view: GEO isn't a separate product. The levers that work are the SEO levers (access, indexing, ranking for the right questions, useful pages) and the PR levers (being mentioned by others). A "GEO package" that sells a file, a schema bundle or a guaranteed ChatGPT spot instead is selling the parts Google told you to ignore. Chapter 9 lists the red flags.

Each layer rests on the one below. Skip a lower layer and nothing above it holds.

Your first check

Before any writing or outreach, make sure AI search bots can read your site at all. A robots.txt rule or a firewall setting can quietly block them. Chapter 6 covers the details; the quick test takes a minute.

Then fix the SEO base, because everything in this guide sits on it. If you haven't read them, Guide 1 chapter 5 explains AI answer engines for beginners, and Guide 3 chapter 8 covers AI search for technical SEOs.

Next, chapter 2 looks at how each AI answer engine actually finds the pages it cites.

If you know where an engine gets its pages, you know which door to knock on. A site that's invisible in Bing will struggle in Copilot. A site that blocks OAI-SearchBot won't be shown in ChatGPT search answers, except as a plain navigational link. This chapter goes engine by engine, with dated numbers, and ends with why you shouldn't treat any of those numbers as fixed.

The common pattern: retrieve, ground, write, cite

Almost every AI answer engine works the same basic way. It doesn't answer from memory alone. It searches, reads what it found, and writes an answer based on those pages. Google's 2026 guide calls this retrieval-augmented generation, "also known as grounding", and says its version relies on "our core Search ranking systems to retrieve relevant, up-to-date web pages from our Search index".^[13]

The second idea is query fan-out: the model turns one question into "a set of concurrent, related queries" to fetch more results.^[13] Someone asks AI Mode "best coaching for NEET in Kanpur". Behind the scenes it might also search for fees, batch timings, results, reviews and distance from a locality. The answer then pulls passages from pages that rank for those sub-searches, as well as for the main one.

One question becomes several searches, and each is another chance to be quoted. Site names are made up.

Mike King of iPullRank explains it with a raffle. Every relevant passage you have is a ticket; more relevant passages across more sub-questions mean more tickets in the draw.^[22]

The query fan-out technique…is the big thing that I think we need to be thinking about in our space.^[22]

King built a tool, Qforia, that simulates fan-out queries for a question. It's a simulation, not Google's real list, so use it for ideas.

Which index feeds which answer

This table is the one to keep. Scale figures are the latest we could source, with dates.

Sources for the table are in each section below. Chapter 6 covers every bot in detail, with robots.txt examples.

Which index feeds which answer, and which bot has to reach your pages.

Google AI Overviews and AI Mode

Google is the simplest case. Its AI features use Google's index, and a page is eligible if it's indexed and can show with a snippet. Nothing else is required.^[14] Google says both features may use query fan-out, so they can show "a wider and more diverse set of helpful links" than classic results.

Pew Research Center's study of real browsing (900 US adults, March 2025) gives a picture of what the summaries cite:^[10]

• 88% of AI summaries cited three or more sources; only 1% cited a single source.
• The median summary was 67 words long.
• Government (.gov) sites were 6% of cited sources, against 2% in standard results.
• Wikipedia, YouTube and Reddit together were 15% of AI summary sources (17% of standard results).

Longer queries trigger summaries far more often. That matters, because buyers' questions are usually long.

Searches that showed an AI summary, by query type

See as a table

Item	Value
1-2 words	8%
Full sentence	36%
10+ words	53%
Question form	60%

Source: Pew Research Center, 68,879 US searches, March 2025^[10]

AI Mode reaches further beyond the normal results. Moz reported on 18 February 2026 that 88% of AI Mode citations were not in the organic results for the same query, across nearly 40,000 US and UK queries.^[24] It's one vendor's data (Moz sells AI tracking), but the direction fits what Google says about fan-out.

Google keeps adding ways for sites to show up inside AI answers. Preferred sources (where users pick sites they like) were extended to AI Mode and AI Overviews on 27 May 2026.^[25] On 15 October 2026 Google announced more and higher inline links in AI answers and a test of an "official websites" section.^[26]

ChatGPT search

OpenAI launched ChatGPT search on 31 October 2024 and opened it to everyone, logged in or not, on 5 February 2025.^[5] It draws on third-party search providers and content from partners.^[27] OpenAI hasn't published a full list of providers.

The bot that matters is OAI-SearchBot. OpenAI's own page says: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers, though can still appear as navigational links."^[28] GPTBot, the training crawler, is a different bot, and blocking it doesn't block search. Chapter 6 has every OpenAI bot in its table.

What does ChatGPT cite? Profound analysed about 680 million citations from August 2024 to June 2025 (it sells AI tracking, so it's a vendor study). For ChatGPT, Wikipedia was far ahead at 7.8% of all citations, then Reddit 1.8%, Forbes 1.1% and G2 1.1%.^[29] Ahrefs found ChatGPT cited the newest pages of the engines it compared.^[30]

ChatGPT is also where most AI referral visits come from: about three in four in SE Ranking's 2026 data (chart in chapter 8).^[11]

Perplexity

Perplexity runs its own crawler. PerplexityBot "is designed to surface and link websites in search results on Perplexity", and Perplexity says it isn't used to crawl content for AI foundation models.^[31] Its user fetcher, Perplexity-User, is covered in chapter 6.

In Profound's data, Perplexity leaned hard on Reddit (6.6% of all citations), then YouTube 2.0%, Gartner 1.0% and Yelp 0.8%.^[29] Its CEO said in June 2025 that Perplexity had handled 780 million queries in May;^[23] later "billion a month" figures are third-party estimates.^[32]

Perplexity also behaved differently in Lily Ray's small study of sites that lost Google traffic in early 2026: it was the one engine where most of them gained citations (chapter 3).^[33]

Gemini

The Gemini app grounds its answers on Google Search, so normal Google SEO applies. The one control to understand is Google-Extended, a robots.txt token that controls Gemini training and grounding in Gemini apps and Vertex AI. It doesn't remove you from AI Overviews or AI Mode (chapter 6).^[34]

Gemini had 950 million monthly active users in July 2026, according to Pichai's earnings remarks.^[8] One vendor report (Tinuiti with Profound) found AI Overviews cite social media more than four times as often as Gemini does.^[35] So even two Google products, on the same index, pick differently.

Microsoft Copilot

Copilot answers draw on Bing's index. Bing's guidelines, rewritten in February 2026, treat grounding and citations in Copilot as outcomes a page can be eligible for. According to Search Engine Journal's report, they advise stating facts directly, keeping entity names consistent and focusing each URL on one topic.^[19] The same report says NOARCHIVE keeps a page out of Copilot responses and grounding results.^[19] IndexNow tells Bing when a page changes, so it can recrawl it sooner.^[36] Chapter 6 covers NOARCHIVE, NOCACHE and the other directives.

Bing's index reaches beyond Copilot. It powers Yahoo's results and most of DuckDuckGo's links.^[37] So if you've never set up Bing Webmaster Tools, do it now. Its AI Performance report, in public preview since 10 February 2026, is the only first-party count of AI citations we've found (chapter 8).^[38]

Claude

Anthropic runs three bots: ClaudeBot for training, Claude-SearchBot for search and Claude-User for fetches a person asks for. All three follow robots.txt.^[39] The one to allow for Claude's search answers is Claude-SearchBot.

Which search index Claude uses isn't stated on Anthropic's crawler page. In March 2025 Simon Willison found Claude's web search results matched Brave Search exactly for his test query. Anthropic's Trust Center listed Brave Search as a subprocessor from 19 March 2025.^[40] Treat that as likely, not confirmed. Claude sent 2.62% of AI referral visits in SE Ranking's 2026 data.^[11]

The engines disagree with each other

Chen, Wang, Chen and Koudas compared AI search engines in a September 2025 paper. They found a "systematic and overwhelming bias towards Earned media" (third-party coverage) over brand-owned and social content, unlike Google's more balanced results. The engines also differed a lot from each other in how many domains they cited, how fresh the sources were, and how much the answer changed with the wording of the question.^[41] Chapter 5 is about earning that coverage.

Here's how different the favourites are, from the same Profound dataset:

Share of all citations, August 2024 to June 2025; vendor study.^[29]

Shares move fast

Every percentage in this chapter is a dated snapshot. In Tinuiti's tracking, social media's share of all AI citations rose every month from October 2025 to January 2026, when it topped 9%.^[35] AI Overviews moved to Gemini 3 in January 2026, according to Ahrefs, and Ahrefs saw the share of citations from top-10 pages change sharply between its two studies.^[43] If someone shows you "the sources ChatGPT trusts" without a date range, ask for one.

What to do with this

Three things follow, and the rest of the guide builds on them:

1. Be in both indexes. Google feeds Google's AI features and Gemini; Bing feeds Copilot. Guide 1 chapter 4 compares the engines.
2. Let the search bots in: Googlebot, Bingbot, OAI-SearchBot, PerplexityBot and Claude-SearchBot. Check with our tool below, then read chapter 6.
3. Rank for the sub-questions as well as the head term. Chapter 3 shows what the research says about ranking and citations.

Most GEO advice you'll read online quotes "up to 40%" from one paper and stops there. This chapter reads that paper properly, puts the main follow-up study next to it, and then looks at what citation data from Ahrefs, Lily Ray, Glenn Gabe and Seer says about ranking, freshness and clicks.

The GEO paper, in plain words

Aggarwal and colleagues built a test bench of about 10,000 queries, called GEO-bench, from nine sources including MS MARCO, Natural Questions, ELI5 and Perplexity's Discover page. About 80% of the queries are informational.^[1]

They didn't test a real search engine at first. They built a simple one:

1. Take the top 5 Google results for a query.
2. Give their full text to GPT-3.5 (gpt-3.5-turbo) and ask it to write an answer with citations, five times per query.
3. Rewrite one of the five sources with a method, such as "add statistics".
4. Measure how much of the answer comes from that source, weighted towards citations that appear earlier. They call this position-adjusted word count.

How the GEO paper (Aggarwal et al., 2023) ran its test: a lab set-up, not a live search engine.

What it found

Here are the nine methods, as relative change against the unchanged source. We computed the percentages from the paper's Table 1 (the overall position-adjusted word count, where the baseline score is 19.3).

GEO paper: change in a source's visibility by method

See as a table

Item	Value
Quotations	40.9%
Statistics	30.6%
Fluency	28.0%
Cite sources	27.5%
Technical terms	17.6%
Easy to read	14.0%
Authoritative	10.4%
Unique words	6.2%
Keyword stuffing	-8.3%

Source: Aggarwal et al., GEO-bench, position-adjusted word count, relative to no change^[1]

They also tried the methods on the real Perplexity. Quotations still came first and keyword stuffing last, but the gains were smaller: quotations +20.7%, statistics +8.7%, and keyword stuffing -9.1% (our calculation from the paper's Perplexity table, Table 5).^[1]

Three findings matter more than the headline:

• Lower-ranked sources gain most. For the source in fifth place, cite sources gave +115.1%, quotations +99.7% and statistics +97.9%. For the source in first place, the same methods lowered its visibility: -30.3%, -22.9% and -20.6%.
• Combining helps a little. Fluency plus statistics beat any single method by more than 5.5%, tested on a 200-example subset.
• Topic changes what works. Statistics helped most for law and government, debate and opinion queries. Quotations helped most for people and society, explanation and history.

Its limits

This is one lab study. The "engine" was GPT-3.5 reading five sources, and only one source was rewritten at a time while the other four stayed as they were. Real engines retrieve from far more pages, use newer models, and change often. The paper tells you how an LLM quotes text it has already been given. It says nothing about whether your page gets retrieved in the first place.

The 2025 follow-up: C-SEO Bench

Haritz Puerto and colleagues tested exactly that gap. C-SEO Bench was accepted at NeurIPS 2025 (Datasets and Benchmarks).^[44] They took ten rewriting methods, eight of them from the GEO paper, and tested them in question answering (web, news, debate) and product recommendation (retail, video games, books). The answering models were gpt-4o-mini, Claude 3.5 Haiku, o3 and o4-mini. Crucially, they also tested what happens when several competing pages use the same tricks.

The authors' summary: most current methods are "largely ineffective", and they often lower the page's ranking. In their full results, only 3 of 54 cases showed a statistically significant gain. Adding statistics lowered rankings in 19 of 24 settings. As more competitors adopted the methods, the gain per adopter shrank towards zero.

What did work? Moving the page higher in the list of documents the model was given. That is their stand-in for traditional SEO, and its effect was much larger than any rewrite. The authors conclude these rewrites should add to SEO, not replace it.

Read together, they say: how you write changes how much of your page gets quoted once it's retrieved. Being retrieved at all is the bigger lever. That's ranking.

Do you need to rank? The citation data

Ahrefs studied 863,000 keyword results and about 4 million AI Overview URLs, published 2 March 2026 (Ahrefs sells an AI tracker, so it's a vendor study).^[43]

Where Google AI Overview citations rank in organic results

See as a table

Item	Value
Top 10	37.9%
Positions 11-100	31.2%
Beyond 100	31.0%

Source: Ahrefs, 863K keyword SERPs, March 2026; its July 2025 study found about 76% from the top 10^[43]

The July 2025 version found about 76% of citations came from the top 10. The two aren't directly comparable: Ahrefs changed its parser, doubled the dataset, and AI Overviews moved to Gemini 3 in January 2026. Ahrefs' best guess is query fan-out: Google now cites pages that rank for the sub-questions, not for the original search. YouTube stood out: it made up 18.2% of cited pages that didn't rank in the top 100, and those YouTube pages alone were 5.6% of all AI Overview citations.

So "you must be top 10" is out of date, but "ranking doesn't matter" is wrong too. You need to rank somewhere in what the engine retrieves, for the question or its sub-questions. Lily Ray's summary from a May 2026 podcast: "it's almost always ranking number one on Google gets you more citations".^[45]

Cited pages rank for the main question or for one of its sub-questions. Names are made up.

Lose Google, lose AI

The strongest real-world evidence runs the other way: when a site drops in Google, its AI citations tend to drop too.

Lily Ray compared 11 subfolders hit by an unconfirmed Google update in mid-January 2026, using Ahrefs data from 20 January and 16 February. Organic traffic fell between 5.7% and 53.1% (26.7% on average). AI citations fell for all 11, by 22.5% on average, and ChatGPT citations by 27.8%. Perplexity was the exception. It's a small sample, and she says so. Her line: "the fastest way to lose visibility in AI search might be to lose it in Google first."^[33]

Glenn Gabe saw the same with NationalToday.com. Its section of more than 850,000 AI-generated local news pages got a scaled content abuse manual action and dropped out of Google's index. Its ChatGPT citations largely disappeared too; he didn't give exact counts.^[47]

Drop in Google and you can drop in AI Search.^[48]

Freshness

Ahrefs looked at about 17 million citations across ChatGPT, Perplexity, Gemini, Copilot, AI Overviews and organic results (July 2025). AI-cited URLs averaged 1,064 days old, against 1,432 days for organic results: 25.7% fresher. ChatGPT cited the newest pages; AI Overviews and organic results the oldest.^[30] That's a correlation from a vendor, not proof that updating gets you cited.

Real updates are worth doing anyway. Fake ones aren't: Google's helpful content guidance says not to change a page's date to look fresh when the content hasn't substantially changed.^[49] Guide 2 chapter 6 covers refreshing content properly.

Being cited and clicks

Seer Interactive measured 3,119 informational queries across 42 client organisations (published November 2025). When a brand was cited in the AI Overview, its organic CTR was 0.70%, against 0.52% when it wasn't cited.^[46] Seer itself warns that strong brands may simply get cited more, so the citation may not cause the clicks.

And many AI sessions never leave the answer. In Kevin Indig's AI Mode usability study of 37 US participants, 77.6% of sessions had no external visit.^[50] Small sample, US only, but it fits Pew's 1% click rate on links inside AI summaries from chapter 1.

How to read a GEO study in 60 seconds

Run through this before you quote any GEO number, including the ones in this guide:

What to do with this

Fix the SEO base first, then check whether your key pages rank for the main question and the questions around it. Search Console's Performance report shows the queries you already appear for, free. For questions you don't appear for yet, a paid tool such as Ahrefs or Semrush shows who ranks. Chapter 4 shows how to list the sub-questions.

One tactic lowered a page's share of the answer both in the lab and on Perplexity: keyword stuffing. If you're unsure whether a page overdoes it, check it. Density is a warning light, not a target.

Chapter 4 turns the useful half of the GEO paper (quotes, numbers, sources, clear writing) into a way of writing pages for people that also happens to suit AI answers. Guide 2 chapter 1 covers keyword research, and Guide 3 chapter 10 covers spam and updates.

Chapter 3 showed that being retrieved matters more than any rewrite. This chapter is about what happens once you are retrieved: whether the engine can find a clean, quotable answer on your page, and whether that answer is worth trusting. The good news is that everything here also makes the page better for the person reading it. Use that as the test: if a change only makes sense for a bot, don't make it.

Why passages, and where Google disagrees

AI answers quote passages, not whole pages. Aleyda Solis built her AI search checklist on this idea: AI systems retrieve chunks of a page, so each passage should make sense on its own.^[51] Mike King gives similar advice: clear sections with descriptive headings, and specific statements written as plain subject-predicate-object facts, which he calls semantic triples.^[22]

Google, though, lists "chunking" content among the things you can ignore, together with rewriting content just for AI systems.^[13] These two positions fit together better than they look. Google is warning against cutting one good page into many thin ones, or splitting text into fragments for machines. Solis and King are describing well-organised writing. Do the second and avoid the first: one strong page per topic, with clear self-contained sections, because people scan too. In Nielsen Norman Group's study, 79% of users scanned a new page and only 16% read word by word.^[52]

The answer-first section

Every chapter of our guides opens the same way: a question, a 40 to 60 word answer, then the detail. Use that pattern for each important section of a page:

1. A heading that names the question in the reader's words ("How much does a root canal cost in Delhi?").
2. The direct answer in the first one or two sentences.
3. The detail: conditions, exceptions, numbers, examples.
4. A link to the source or to the next step.

Why 40 to 60 words? Pew found the median Google AI summary was 67 words long (March 2025).^[10] That's our inference, not a rule anyone has published: a short, complete answer is easy to quote inside an answer of that size, and a reader can use it too.

Six parts that make a passage easy for a reader, and an AI answer, to quote.

Numbers, quotes and sources

In the GEO paper, adding quotations, statistics and cited sources raised a source's share of the answer by 27.5% to 40.9% in the lab, mostly for sources ranked lower down.^[1] Chapter 3 explained why that result is weaker in competitive tests. Do it anyway, for a simpler reason: a number with a source is checkable, and a page people can check is a page they trust.

Our rules, the same ones this guide follows:

• Every number has a source, a sample and a date. "AI summaries cite three or more sources 88% of the time (Pew, 900 US adults, March 2025)" beats "most AI answers cite several sources".
• Quotes are word for word, from a named person, with a link. If you can't find the exact words, paraphrase without quote marks.
• Link to the primary source, not a blog that quoted it.
• Say what the number doesn't cover. US data is US data.

Here's a section from a made-up agency page, before and after:

The second version answers the heading's question in its first word, gives one sourced number with its sample and date, and points to Google's own words. Any AI engine, or any reader, can lift it out and use it.

Non-commodity content

Google's 2026 guide asks for "non-commodity" content. Its example contrasts "7 Tips for First-Time Homebuyers", which anyone could write, with a homeowner's first-hand account of inspecting a sewer line.^[13] John Mueller said it a year earlier:

Focus on making unique, non-commodity content that visitors from Search and your own readers will find helpful and satisfying.^[53]

For an Indian business, non-commodity means the things only you have. A coaching institute's own pass rates by year. A clinic's real price range for a procedure. A D2C brand's photos of how a product is made. A SaaS company's numbers from its own users. An AI engine can find "10 tips for NEET preparation" on a thousand pages. It can find your institute's last three years of results only on yours.

Cover the sub-questions, on one page

Query fan-out means AI answers pull from pages that rank for the questions around the main one (chapter 2). Solis's advice is to cover every side of a topic in depth.^[51]

The trap is making a page for every variation. Google is direct about it: creating separate content for every way people might search, mainly to manipulate rankings or generative AI responses, violates its scaled content abuse policy.^[13] So the rule is one strong page per intent, with a section for each sub-question. "Root canal cost in Delhi" gets one page, with sections on what changes the price, insurance, the number of visits and aftercare. Not five pages.

To list the sub-questions, look at Google's "People also ask" boxes, your Search Console queries, and the questions customers ask on calls. AnswerThePublic gives about three free searches a day.^[54] Our free Question Finder collects them from autocomplete, merges the ones you copy from the box, and exports the heading list.

Tables, images and video

Make comparisons real HTML tables, not images of tables, so any system can read them. Give images descriptive alt text and captions.^[51] Google's May 2025 advice for AI search also says to go beyond text, with images and video, and to keep Merchant Center and Business Profile up to date.^[53]

Video deserves a mention of its own. In Ahrefs' March 2026 study, YouTube pages that didn't even rank in the top 100 for the query made up 5.6% of all AI Overview citations.^[43] If your competitors only write, a short video that answers the same question is an opening.

Authorship and dates

Google's helpful content page asks you to make clear who created the content, how, and why. That means bylines where readers expect them, and an honest note on how it was made, including AI use where readers would wonder.^[49] Solis's checklist describes citation-worthy content as accurate, current, verifiable, sourced and clearly authored.^[51]

In practice: a byline linked to a real author page, a reviewer for medical, legal or financial topics, and a "last updated" date that changes only when the content does. Guide 2 chapter 5 covers E-E-A-T in more depth.

What not to do

The mass-AI-pages warning has numbers behind it. Lily Ray tracked more than 220 sites publicly named as customers of AI content platforms, using Ahrefs and Sistrix estimates up to May 2026. 54% had lost 30% or more of their peak organic traffic, and 22% had lost 75% or more. Her summary: "it works, until it doesn't."^[55] It's a correlation, not a proven cause. The typical shape she saw was fast growth for six to twelve months, a peak, then a drop that erased most of the gains, often to below where the site started.

Evidence vs hype: our verdict on common tactics

If you use AI to help write

Plenty of teams draft with ChatGPT or Gemini. That's allowed. Google's guidance, last updated on 1 October 2026, is plain about checking the output:

Our rule is simple. AI can help with outlines and first drafts. It never supplies a number, a quote or a date; those come from the source, checked by a person. Tools such as Surfer, Clearscope and Frase are fine for building a brief. Don't write to hit their score.

Check the result with our tools: keyword density as a warning light, schema that matches the visible page, and titles and descriptions that still say what the page does.

Guide 2 chapter 4 covers content briefs. If you'd like us to rewrite your key pages this way, our content marketing service does exactly that. Next, chapter 5 moves off your site: what other sites say about you.

Chapters 3 and 4 were about your own pages. This one is about everyone else's. When an AI engine recommends "the best physiotherapy clinics in Pune" or "a good invoicing tool for small businesses", it's summarising what the web says. If the web says little about you, there's little to summarise.

What correlates with AI visibility

The most-quoted study here is from Ahrefs. Louise Linehan looked at 75,000 brands (sites with Domain Rating above 40) and measured which factors went together with being mentioned in Google's AI Overviews. Published May 2025, it's a vendor study, and it measures correlation (Spearman), not cause.^[57]

Brand factors vs mentions in AI Overviews

See as a table

Item	Value
Web mentions	0.664
Branded anchors	0.527
Branded searches	0.392
Domain Rating	0.326
Backlinks	0.218

Source: Ahrefs, 75,000 brands, Spearman correlation, May 2025. Correlation, not cause^[57]

Branded web mentions, meaning your brand named on other sites with or without a link, correlated far more strongly than backlinks did. About 26% of the brands had no AI Overview mentions at all. Ahrefs' own summary is that AI visibility depends on how widely your brand appears across the web, not just on your website.

Correlation can mislead in both directions here. Big brands get mentioned more and get cited more, and both may come from being big. Still, it points the same way as the academic work below.

Earned media bias

Chen and colleagues' September 2025 paper compared AI search engines with Google. They found the AI engines had a "systematic and overwhelming bias towards Earned media": third-party coverage such as reviews, press and independent sites, over a brand's own pages and its social posts. Google's results were more balanced. They also warn of a "big brand bias" that puts niche players at a disadvantage.^[41]

For a small Indian brand, that's both the problem and the plan. You won't out-shout the big names on your own blog. You can get written about in the places that cover your niche.

AI engines lean on what others write about you (Chen et al., 2025; Profound).

Where AI looks

Each engine has favourite third-party sources (chapter 2 has the full table). In Profound's data from August 2024 to June 2025, ChatGPT cited Wikipedia most; AI Overviews and Perplexity cited Reddit and YouTube most.^[29] Pew found Wikipedia, YouTube and Reddit together made up 15% of the sources in Google's AI summaries (US, March 2025).^[10] A Tinuiti and Profound report put social media at about 9% of all AI citations in January 2026.^[35]

We looked for India-specific data on which sites AI engines cite for Indian queries, or on AI referral traffic in India. We didn't find any published study. So the platforms above are a starting point to test with your own prompts, not a rule for Indian buyers.

Entities: one set of facts everywhere

An entity is a thing a search engine can identify: your business, its founder, its products, its address. Google's Knowledge Graph stores entities, and knowledge panels, the boxes that appear for some brands and people, are generated automatically from information across the web.^[58] If you're the official representative of an entity that has a panel, you can claim it and suggest changes. The help page doesn't describe any way to request one.

What you control is consistency. Your name, address, phone, founders, services and prices should match on your site, Google Business Profile, directories, marketplaces and social profiles. Organization and Profile page markup are among the structured data types Google supports, so mark up your About and author pages, and make sure the markup matches what the page shows.^[59] Guide 3 chapter 7 goes deeper on entity SEO.

Jono Alderson made the point well when writing about AI agents that check prices and policies on your behalf:

Making three contradictory prices machine-readable gives the agent a better-documented problem.^[60]

A coaching centre with one fee on its website, another on Justdial and a third in a Facebook post has that problem. Fix the facts before you add markup.

One Organization, joined to its people, place, pages and profiles. Names are made up.

Earning mentions that count

Mentions that count are ones a journalist, reviewer or community member chose to write. A few ways to earn them:

• Publish your own data. A survey of your customers, prices you track, results from your own work. Data gives writers a reason to name you. Guide 2 chapter 9 covers digital PR.
• Trade press and local news. An expert comment, a case, a useful number. Indian-language news sites count too: Techmagnate, by its own account, used PR on Hindi news sites as part of a vernacular SEO campaign for Bajaj Finserv.^[62]
• Reviews. Ask real customers, on the platforms your buyers read.
• The same message in several places. Mike King points out that the passages AI answers use can come from other sites as well as yours: LinkedIn, Reddit, YouTube, review sites, news and trade sites.^[22] Keep your key facts the same on all of them.

Kevin Indig and Amanda Johnson suggest giving off-site distribution and community time their own budget line, with a named owner, and spending about 5% of the marketing budget on community (10% in some cases).^[63] That's their suggestion for planning, not a measured benchmark, but it's a useful way to stop mentions being nobody's job.

That story is about links, mentions and traffic, not AI citations. The point is the mechanism: give people something worth citing, and mentions can follow.

Communities without spam

Reddit, Quora, YouTube comments, Facebook groups and LinkedIn are where buyers ask each other for recommendations, and where AI engines often look. Take part as a named person who works for the business, say so, and answer the question asked. Don't drop links into every thread.

This is where GEO most often turns into spam. Google lists "seeking inauthentic mentions" as something to ignore for its AI features,^[13] and its spam policies cover both user-generated spam and attempts to manipulate generative AI responses.^[15]

The platforms have their own rules. Wikipedia requires anyone paid to edit to disclose their employer, client and affiliation. It strongly discourages paid editors from editing affected articles directly; they should propose changes on the article's talk page instead.^[64] Reddit's own rules ask you to "participate authentically in communities where you have a personal interest" and not to spam,^[65] and every community (subreddit) adds its own rules on self-promotion, so read them before you post. Many ban it outright.

What stopped working

For a while, a popular GEO trick was publishing self-promotional "best X" listicles that ranked your own product first, on your site and others. Lily Ray says these worked for a time and are now treated as spam and risky.^[45] In her tracking, at least 40 sites with many self-promotional listicles fell 40% to 95% between January and April 2026.^[55]

Local businesses

For clinics, salons, coaching centres and shops, the most important third-party profile is your Google Business Profile. Google's May 2025 advice for AI search includes keeping Business Profile and Merchant Center information up to date.^[53] Guide 2 chapter 10 covers local SEO step by step.

Use of AI for local choices is growing fast, at least in the US. BrightLocal's 2026 survey of 1,002 US adults found 45% had used ChatGPT or another AI tool for local business recommendations, up from 6% in 2025.^[66] That's a vendor survey of US consumers. We found no Indian equivalent.

Where to be mentioned, by business type

These are our suggestions from the citation data above; test them with your own prompts.

Brand is the long game

Rand Fishkin has argued that the main shift is zero-click search, with AI as a secondary part of it. His answer is short:

Brand is the solution.^[67]

Seer's CTR data in chapter 3 points the same way, with the same caveat that strong brands may simply be cited more.^[46] Either way, the brands people already know and talk about are the ones AI engines name.

Tools for this chapter

Check that your Organization or LocalBusiness markup matches the page, and that your Open Graph tags show the same name and description when people share your links.

We don't have a backlink or mention checker yet. Meanwhile, Search Console's Links report shows who links to you, its branded queries filter shows whether people search for your name more over time, and Ahrefs Free shows backlinks to your own verified sites. For mention tracking at scale, Ahrefs Brand Radar AI starts at $199 a month (price seen 16 October 2026).^[68]

Chapter 8 shows how to measure whether any of this changes your AI visibility. If you want help earning coverage, that's part of our content marketing service.

Everything in chapters 2 to 5 assumes one thing: the engine can actually read your page. This chapter makes sure it can. It's the full version of the short crawler section in Guide 3, chapter 8, and it's the one technical job in GEO where a single wrong line can switch you off.

Three kinds of AI bot

Most people talk about "AI crawlers" as one thing. They're not. Each company runs bots with different jobs, and the job decides what happens when you block one.

• Training crawlers collect pages that may be used to train future models. Blocking them is an opt-out from training. It doesn't stop the company's search product from finding you.
• Search crawlers build the index an AI search product looks things up in. Block one and you drop out of that product's answers.
• User fetchers open a single page because a person asked the assistant about it. They don't crawl the web on their own, and some of them don't follow robots.txt at all.

The bot's job decides what blocking it does. Names as vendors list them, October 2026.

The AI bot table, by company and purpose

Every row below comes from the company's own crawler page, read on 16 October 2026. Bot names change, so check the vendor pages again before you edit your file.

Two details are easy to miss. Apple's normal crawler, Applebot, may itself help train Apple's models as well as feed its search features, so Applebot-Extended is the opt-out for the training use only.^[70] And Amazonbot and Bytespider (ByteDance) show up in server logs in large numbers. This guide doesn't cover their vendors' pages, so read those before you write rules for them.

robots.txt by purpose

Here's the pattern we suggest: say no to training if you want to, and say yes to every search bot.

User-agent: GPTBot
Disallow: /

User-agent: Google-Extended
Disallow: /

User-agent: OAI-SearchBot
Allow: /

User-agent: PerplexityBot
Allow: /

Add ClaudeBot and CCBot to the disallowed groups and Claude-SearchBot to the allowed ones if you want the same split for Anthropic and Common Crawl. If you're happy for your pages to be used in training, leave the training groups out. No rule means allowed. One trade-off: blocking Google-Extended also stops the Gemini app grounding its answers on your pages, so leave that group out if Gemini matters to you.

Say no to training if you want to, and yes to every search bot.

One rule from Google's robots.txt specification catches people out more than any other here. A crawler obeys only the group with the most specific user agent that matches it, and named groups aren't combined with the * group.^[73] So if you give GPTBot its own group, your User-agent: * rules no longer apply to it. Repeat any shared Disallow lines inside the named group.

The other limits (the 500 KiB cap, caching for up to 24 hours, what a 5xx on robots.txt does) are in Guide 3, chapter 8, with the edge cases in Guide 3, chapter 1. Guide 1, chapter 8 covers the basics. Test after every change, not once a year.

Google's own controls for AI features

For AI Overviews and AI Mode, Google uses the same controls as the rest of Search: robots.txt rules for Googlebot, and the nosnippet, data-nosnippet, max-snippet and noindex directives.^[14] A page has to be indexed and eligible for a snippet to appear at all. So a stray nosnippet doesn't just shorten your result. It can keep the page out of AI answers.

Since 31 August 2026 every site also has a dedicated switch. In Search Console, go to Settings, then Search generative AI. "Include" is the default. "Exclude" keeps your links and content out of AI Overviews, AI Mode and generative AI features in Discover, and you get no impressions or traffic from them.^[17] Google says the setting isn't a ranking or inclusion signal for the rest of Search, it doesn't affect AI training (that's Google-Extended), and changes generally take a few days to apply.

Our view: very few businesses should flip this to Exclude. It's there for publishers who've decided the trade isn't worth it. Know it exists, and check nobody on your team switched it by mistake.

Bing and Copilot

Microsoft's documentation for Microsoft 365 Copilot Chat says Copilot writes a short web query from your prompt and sends it to Bing.^[69] So Copilot visibility is Bing visibility: let Bingbot in and set up Bing Webmaster Tools (Guide 1, chapter 8).

Bing's rewritten guidelines also explain how its directives work in AI answers. According to Search Engine Journal's report of February 2026, NOARCHIVE keeps content out of Copilot responses and grounding results, NOCACHE limits Copilot to the URL, title and snippet, and nosnippet can lower citation quality.^[19] If a developer once added NOCACHE or NOARCHIVE "for privacy", that's a reason Copilot cites you thinly or not at all.

To keep Bing current, use IndexNow. It tells Bing, Naver, Seznam.cz, Yandex and Yep that a URL changed, so they can recrawl it sooner. Google doesn't use it.^[36]

User-triggered fetchers, and how to tell a real bot from a fake

ChatGPT-User, Perplexity-User and Meta's Meta-ExternalFetcher may ignore robots.txt, and their makers say why: a person asked for that page, so the fetch is treated like a visit, not a crawl.^[71] Anthropic is the exception here and says Claude-User follows robots.txt. If you really need to stop user fetchers, robots.txt won't do it.

Anyone can type "GPTBot" into a user agent string, so don't trust the name alone. The vendors publish their IP address lists:

• OpenAI: openai.com/gptbot.json, openai.com/searchbot.json and openai.com/chatgpt-user.json^[28]
• Perplexity: perplexity.com/perplexitybot.json and perplexity.com/perplexity-user.json
• Anthropic: claude.com/crawling/bots.json^[39]
• Google: developers.google.com/crawling/ipranges/^[74]
• Common Crawl: a JSON list of CCBot's ranges, and it suggests a reverse DNS check, since fake CCBots exist

Compare the IP in your server log with the list. We don't have a DNS or IP lookup tool yet, so use your host's log viewer or the host command on a Mac or Linux terminal for now. One warning from Anthropic: don't block its bots by IP, because then they can't read your robots.txt and your opt-out may not work.^[39]

What the crawling costs you, and what it gives back

Cloudflare compared how often each company's bots crawl with how many visits its products send back. These are July 2025 figures from Cloudflare's network.^[75]

The same post found that 79% of AI crawling in July 2025 was for training, 17% for search and 3.2% for user actions. The ratios move fast: Anthropic's was about 286,930 in January 2025.

Share of AI-only crawling, July 2025

See as a table

Item	Value
GPTBot	28.1%
ClaudeBot	23.3%
Meta-ExternalAgent	17.7%
Amazonbot	14.1%
Bytespider	5.8%

Source: Cloudflare, "The crawl-to-click gap", August 2025, Cloudflare network data^[75]

So should you block training bots? Decide by business, not by fear. A news publisher or a paid course provider whose content is the product has good reasons to say no to training. A dental clinic in Pune or a SaaS company in Bengaluru usually wants its name, prices and services known as widely as possible, and gains little by blocking. Either way, keep the search bots open.

Rendering: can AI crawlers read JavaScript?

Google renders pages with a recent version of Chrome,^[76] and Googlebot fetches up to 2MB of each HTML file. Anything past that isn't rendered or indexed.^[77]

Most AI crawlers are less patient. In Vercel and MERJ's study of traffic across Vercel's network, published in December 2024, GPTBot, ClaudeBot and PerplexityBot didn't render JavaScript. ChatGPT's and Claude's crawlers fetched JavaScript files but didn't run them. Applebot did render, through a browser-based crawler, and Gemini uses Google's infrastructure.^[78] That's one network and one point in time, and the bots may have changed since. We haven't found a newer first-party statement from OpenAI, Anthropic or Perplexity on rendering.

The safe rule: your main text, headings, prices and internal links should be in the HTML the server sends, before any script runs. That's what Aleyda Solis's AI search checklist asks for too: make content reachable and renderable by search and AI crawlers, with server-side rendering and no blocking.^[51] To check a page, open it in Chrome, turn off JavaScript in DevTools and reload. Whatever disappears is what a non-rendering bot can't see. For Google's own rendered view, use URL Inspection in Search Console. Guide 3, chapter 2 goes deeper on JavaScript SEO. A rendered-vs-raw HTML checker is on our list of tools to build.

Don't block what you want cited

Asked about llms.txt, John Mueller said the bigger priority is simple: "don't block agents".^[79] Many blocks aren't decisions at all. They're accidents:

• A "block AI bots" or "bot fight" setting switched on in a CDN, firewall or security plugin, often without anyone remembering
• A staging site's Disallow: / copied to the live site at launch
• A firewall that answers bots with a challenge page or a 403 instead of the page
• A stray noindex, nosnippet or X-Robots-Tag header on templates you want cited
• A slow server that times out before the bot gets the page

robots.txt can say "allowed" while the firewall says no, so check the file and also what the server sends back. Our HTTP Status Checker shows the status code a URL returns, and the Noindex Tag Checker catches stray noindex tags. The status checker sends its own user agent, and you can't pick another. A firewall that blocks only AI bots can still give it a 200, so check your CDN or firewall's bot settings and logs too.

How our checker scores this: its score counts only the search bots and user fetchers you allow, because those decide whether AI answers can cite you. Training bots are shown separately as your choice, and llms.txt is reported for information only, with no points. So blocking GPTBot, ClaudeBot or Google-Extended on purpose doesn't lower your score.

Agents are next

The newest visitors aren't crawlers or chat fetchers but agents that act for a person: compare plans, fill a cart, book a slot. Google's AI optimisation guide points site owners to agent-friendly practices and says protocols like the Universal Commerce Protocol "are emerging that will allow Search agents to do more".^[13] The basics are the same as this chapter: don't block them, and serve real HTML with accurate prices and stock. Guide 4, chapter 9 looks at where this is heading.

With access sorted, the next question everyone asks is whether you also need an llms.txt file. Chapter 7 answers it in five minutes.

Few GEO topics get more questions per word of evidence than this one. Here's the whole picture in five minutes, both sides, so you can decide and move on.

What llms.txt is

Jeremy Howard of Answer.AI proposed llms.txt on 3 September 2024. The idea: a Markdown file at /llms.txt that gives AI tools a short, clean map of a site. Then an assistant or agent working at question time doesn't have to wade through menus, scripts and ads.^[80] The proposal also suggests offering plain Markdown copies of pages.

The format, in order:

1. An H1 with the site or project name. This is the only required part.
2. A blockquote with a short summary.
3. Optional notes: paragraphs or lists, but no headings.
4. H2 sections, each a list of links written as [name](url), with an optional note after a colon.
5. By convention, a section called "Optional" for links an agent can skip when it's short of space.

A minimal file for an Indian coaching institute might look like this:

# Example Coaching Academy

> NEET and JEE coaching in Kota with classroom and online batches. Fees, batch dates and results are on the pages below.

## Courses
- [NEET two-year programme](https://example.in/neet-two-year): Syllabus, timetable and fees
- [JEE crash course](https://example.in/jee-crash-course): 12-week batch, starts every January

## Optional
- [Blog](https://example.in/blog): Exam tips and toppers' interviews

Only the H1 is required. Google Search ignores the file.

We don't have an llms.txt generator yet. Write the file by hand from the llmstxt.org format; for a small site it's a short job.

Who publishes one

Adoption depends heavily on whose sites you count.

Originality.ai's tracked sites put adoption at about 1%. Ahrefs' 28% comes from its own analytics customers, who are more technical than the average site, and Ahrefs itself calls the figure an upper bound.^[82]

Who actually reads it

This is the part that matters. Ahrefs looked at the server logs of those 137,210 domains for May 2026, published on 15 June 2026. About 1,100 files got any requests at all, roughly 22,000 between them, and 96% of those requests came from bots.^[82] Here's who sent them:

Who requested llms.txt files, May 2026

See as a table

Item	Value
SEO audit tools	21.7%
AI bots, all types	19.5%
Unidentified bots	14.9%
Web crawlers	13.1%
Tech profilers	11.6%

Source: Ahrefs, server logs of 137,210 domains, share of all requests to llms.txt^[82]

Inside that 19.5%, AI agents made 10.5% of requests and training crawlers 5.3% (GPTBot 4.51%, ClaudeBot 0.8%). The bots that fetch pages to answer questions, the ones you'd want reading it if it helped citations, made 3.6%: retrieval bots such as OAI-SearchBot and PerplexityBot 1.1%, and assistants such as ChatGPT-User 2.5%. And a request doesn't prove anything was read or used. Ahrefs treats every consumption figure as a ceiling.

The study didn't check whether files followed the format, and it covered only the llms.txt file itself, not Markdown copies of pages. Their own one-line summary: publish one today, and the most likely outcome by far is that nothing ever fetches it.

What Google says

Google has been unusually direct. Its AI optimisation guide lists llms.txt under things you can ignore. You don't need new machine-readable files, AI text files or Markdown to appear in Google Search, including its AI features. Creating them "will neither harm nor help your site's visibility or rankings in Google Search, as Google Search ignores them."^[13] A changelog note on 15 June 2026 added that "it's fine if you want to maintain these files for other services or systems that use them."^[25]

John Mueller wrote in June 2025 that no AI system currently uses llms.txt.^[83] A year later, answering questions on Reddit, he hadn't changed his mind.^[79]

it's purely speculative for now^[79]

His test for when to make one is practical: "When an AI platform that brings you clients complains that it needs the file for your site." Patrick Stox of Ahrefs put it more bluntly in a 2025 conference talk: llms.txt "Isn't used, probably won't be."^[2]

When it does make sense

The case for llms.txt is strongest where the reader is a program, not a search engine. Documentation sites, APIs and developer tools are the clearest fit: an AI coding assistant or agent may be pointed at the file to find the right docs fast. The llmstxt.org site says OpenAI, Anthropic and Google's Gemini publish llms.txt files for their developer documentation, that documentation platforms generate them automatically, and that Chrome's Lighthouse checks for one in its agentic browsing audits.^[80] Those are the author's claims. Ahrefs counted about 22 Lighthouse requests in its whole sample.^[82]

The red flag

If an agency's GEO package leads with "we'll set up your llms.txt", ask what evidence they have that it changes citations. There isn't any public evidence yet. Adding the file is fine. Paying for it as a visibility service isn't.

The real priority is still the one from chapter 6: make sure the search bots can get in.

Our AI Crawler Checker reports whether you have llms.txt and llms-full.txt for information only. It gives no points for them, for the reasons in this chapter. The bot access results are the part that matters.

Next, chapter 8 shows how to measure whether any of this is working.

The obvious follow-up to "can you get us into ChatGPT?" is "how will we know?" There's no single dashboard for it. There are two free first-party reports, one analytics setup and one spreadsheet, and together they give an honest picture.

Where to measure, and what each place can't tell you

Before you read any AI number, run the AI Crawler Checker from chapter 6. A site that blocks OAI-SearchBot will show zero ChatGPT citations, and no report will tell you why.

Google Search Console

Clicks and impressions from AI Overviews and AI Mode are counted in the normal Performance report under the "Web" search type, mixed in with everything else.^[14] You can't filter them out there.

On 3 June 2026 Google added separate generative AI performance reports, one for Search and one for Discover, and had rolled them out to all sites worldwide by 31 August 2026.^[84] They show:

• Impressions: how often your URLs appeared in generative AI features in Search and Discover
• Pages: which URLs appeared
• Countries
• Devices (for Search only)
• Dates by hour, day, week or month

No clicks, no queries, no position. The same data still counts in the overall Performance report; the new reports are just a separate view. Google says it may add more metrics over time.

Drawn by us, not a screenshot. All figures are made up.

Use it to see which pages AI answers draw on. If your clinic's "root canal cost" page gets AI impressions and your "root canal procedure" page gets none, that's a lead worth following. Don't use it to judge traffic. One more thing to check while you're there: if someone set Settings, Search generative AI, to Exclude, these impressions drop to zero by design (chapter 6).^[17]

Two older Search Console features help here. The branded queries filter, open to all eligible sites since 11 March 2026 (whole-site properties with enough impressions, so small sites may not see it), separates people searching your name from everyone else. Rising branded demand is one of the few signs that AI mentions are working.^[85] And custom annotations let you mark the day you changed robots.txt or rewrote five pages, so you can line changes up with results later.^[86]

Finding AI Mode-style queries

Search Console doesn't label AI Mode queries, but you can make a good guess. Glenn Gabe's method: export all queries through the Search Analytics API to get past the 1,000-row limit. Have an AI assistant flag the long, conversational ones and the follow-ups, then look at their clicks. He found few or none on many of them. Check the flagged list by hand before you use it; his included queries from automated tracking tools.^[87] Guide 2, chapter 11 walks through the steps. Treat the result as a list of topics people explore in AI Mode, not as traffic you're missing.

Bing Webmaster Tools: AI Performance

This is the only first-party report we know of that counts citations. Bing launched it in public preview on 10 February 2026. It covers Microsoft Copilot, AI-generated summaries in Bing and some partner integrations, and shows:^[38]

• Total citations: how many times your content was shown as a source in AI answers. It doesn't say where in the answer.
• Average cited pages: the average number of your pages cited per day
• Grounding queries: a sample of the phrases the AI used when it retrieved your content
• Page-level citations and a trend over time

On 16 June 2026 Bing added four views, still in preview. They are Intents (informational, commercial, local and so on), Topics, Citation Share (your share of all citations for a grounding query) and Compare (this period against the last).^[88] Bing is careful about that third one: "Citation Share is designed as an observational metric – not a ranking system or a competitive scoreboard."

Drawn by us, not a screenshot. All figures are made up.

The grounding queries are the most useful part. They're the words an AI system used to find you, and they often read like the questions your customers ask on the phone. Neither Bing announcement mentions clicks, and the report doesn't cover ChatGPT, Perplexity or Gemini.

GA4: AI referrals

When someone clicks a link in ChatGPT, Perplexity, Gemini, Copilot or Claude, the visit usually arrives with the assistant as its source. Since 13 May 2026, GA4's default channel group has its own "AI Assistant" channel^[89] for these visits, from sources like ChatGPT, Gemini, DeepSeek, Copilot and Grok.^[90] Google's AI Overviews and AI Mode aren't in it. Check that channel first. Build the channel below only if you want to add sources Google's list misses.

.*(chatgpt|openai|perplexity|gemini\.google|copilot|claude\.ai|deepseek|grok).*

A standard GA4 property allows two custom channel groups besides the default, so don't waste them on experiments.^[91] If you'd rather not touch the settings, an Exploration with the same regex as a filter on session source gives the same numbers. Ahrefs Web Analytics, included in the free Ahrefs plan for verified sites, also shows traffic from AI assistants.^[92]

One caution, from how referrers work rather than from a study: some AI visits arrive without a referrer, for example from a copied link or an app, and land in Direct. Your AI channel is a floor, not the full number.

Drawn by us, not a screenshot. All figures are made up.

Keep the expectations real

SE Ranking looked at 101,574 sites connected to Google Analytics. AI assistants sent 0.32% of their traffic from January to April 2026, against 42.75% from organic search.^[11] Most of that 0.32% came from one source:

Where AI referral visits came from, 2026

See as a table

Item	Value
ChatGPT	74.78%
Gemini	11.56%
Perplexity	7.23%
Copilot	3.51%
Claude	2.62%

Source: SE Ranking, 101,574 sites, share of AI referral traffic, January to April 2026^[11]

Small doesn't mean worthless. At Ahrefs, Patrick Stox reported that 0.5% of visitors came from AI and drove 12.1% of sign-ups.^[2] That's one software company, not a rule. You'll also see "AI traffic converts 4.4 times better" quoted everywhere; it's a Semrush figure, reported on MarTech (a Semrush-owned site) without any method.^[93] Measure your own key events in GA4 instead.

The manual prompt log

Reports show what Google and Bing count. They don't show whether ChatGPT names you when a parent in Jaipur asks for the best IIT coaching online. For that you ask, by hand, every month.

Run the same prompts every month and read the trend. Names are made up.

Why only a snapshot? AI answers are generated fresh each time, so the same prompt can give different sources an hour apart. And the mix of sources each engine cites shifts within weeks (chapter 2). We've also found no public source for how often people ask a given prompt, so ignore any tool that quotes "prompt volume" as if it were Keyword Planner data.

Paid trackers, honestly

Paid tools run hundreds of prompts on a schedule and chart the results. Here's what the main ones charge for AI visibility, in US dollars, as shown on their own pricing pages on 16 October 2026. We have no affiliate links with any of them.

Remember what Google says about all of them: "No third-party tool has access to our internal ranking or AI systems."^[13] They ask the engines questions, just as your prompt log does, only more often.

Which one when: start with Search Console, Bing Webmaster Tools, GA4 and your own log. They're free and cover most small and mid-sized businesses. Buy a tracker when a client pays for monthly AI visibility reporting across dozens of prompts, or when you track several brands and the manual log takes more than a day.

What to report each month

Kevin Indig and Amanda Johnson suggest reporting three groups together every month: exposure (visibility, mentions, citations), behaviour (branded searches, direct visits) and outcomes (pipeline, conversions).^[63] For AI visibility that becomes:

• Exposure: Search Console AI impressions, Bing citations, prompts where you're cited or named
• Behaviour: branded queries in Search Console, direct and AI-referral visits in GA4
• Outcomes: leads, calls, WhatsApp enquiries and sales, including from AI referrals

Clicks are scarce. Influence is plentiful.^[50]

Put it on one page, with last month beside each number, and build it once in Google's free reporting tool with the Search Console and GA4 connectors (Guide 2, chapter 11). If you run this reporting for clients as a service, Guide 4, chapter 5 covers how to package it.

If you add links to your site on profiles that AI answers often cite, such as LinkedIn or directory listings, tag them with UTM parameters so the visits show up clearly in GA4. Our free UTM Builder makes them, with a LinkedIn preset and the GA4 channel shown for each link.

Take a baseline with the audit before your first monthly report. Chapter 9 turns everything in this guide into a checklist and a 30-day plan.

This chapter is the one to print. It turns the guide into 20 checks, a four-week plan you can start on Monday, and a list of claims you can stop worrying about.

The GEO checklist

Twenty checks in five groups. Each one points to the chapter that explains it. The checklist PDF that comes with this guide has all twenty, plus a few more.

A 30-day plan

This is our plan, built from the chapters of this guide, not from anyone's study. It assumes one person with a few hours a week and a site that's already indexed.

A 30-day GEO plan for one person with a few hours a week.

After day 30, keep the rhythm monthly: five more pages, one more mention, one report. Don't expect the week 4 numbers to move much. Week 4 is about having a baseline and a habit; the changes from weeks 2 and 3 take longer to show in search and AI answers.

Myths to drop

These claims turn up in GEO advice and sales pitches. Here's what the sources actually say, as of October 2026.

GEO myths, checked against the sources, October 2026.

How to judge a GEO vendor

Clients now ask agencies and freelancers "can you get us into ChatGPT?", and some answer with promises no one can keep. Google's guidance on third-party SEO and "AEO/GEO" services says plainly that third-party tools don't have access to its internal ranking data.^[3] Use that as your starting point.

Red flags:

• Guaranteed placement in ChatGPT, Perplexity or AI Overviews. AI answers are generated fresh each time, and no outside tool can see inside the engines.
• Hidden text for AI, or lines meant to instruct a chatbot
• Hundreds of near-identical pages, one for every question variant
• An llms.txt file sold as the main GEO deliverable
• Claims to use "internal" Google data or a secret AI ranking score
• Paid "mentions" placed in bulk on forums or listicles

Questions to ask before you sign:

1. Which engines will you track, and which prompts? Can I see the list?
2. How do you measure, and will I get the raw prompt log and screenshots?
3. What exactly will you change on my site?
4. What will you do off the site, and how does it stay within Google's and Bing's spam rules?
5. What do you report each month besides citations: branded searches, leads, sales?

Do not implement risky and spammy tactics just to rank in AI search.^[47]

He wrote that after a site's AI-generated section was hit by a Google manual action for scaled content abuse, and its ChatGPT citations for that section largely disappeared too. Whatever loses Google's trust tends to lose AI visibility with it.

If you'd like us to run the 30-day plan with you, see our SEO service and content marketing service, or get a quote. Otherwise, start week 1 today, and take the 10-question quiz below to check what stuck.

Test yourself

Ten questions on the whole guide. Pick an answer, check it, and see your score at the end.

Frequently asked questions

Sources