Ask ChatGPT, Claude, Gemini or Google's AI Overviews (the written answer at the top of some Google searches) for a recommendation and you get one answer with a handful of sources. Being one of those sources is the new goal, and an industry of advice has grown up around it. Much of that advice rests on opinion or vendor claims.
We read the platform documentation and the published studies and ranked 22 common tactics by how well the evidence says each one works. "Works" means one thing here: the tactic raises the odds that an AI assistant will find your page, cite it or recommend your product.
The short version: let AI tools reach and read your pages, put real answers and facts only you have on them, earn mentions on credible sites, and skip llms.txt, a guide file some sites publish for AI tools.
The evidence is uneven. The top three ranks rest mainly on platform documentation and crawler logs. Below that, much of it is correlation (two things that rise together, which does not prove one causes the other), lab tests, or studies by firms that sell SEO or AI-tracking tools. A source marked (vendor) is a company that sells tools or services in this field, so it has a stake in the answer.
Each entry rates our confidence in its rank: high, medium or low. Low means the evidence is thin, not that the tactic is bad. One popular tactic has been checked several times in crawler logs and citation data without finding a benefit. It comes last.
Short on time? Read the table below, then jump to the six steps at the end.
The 22 at a glance
Tier 1: open the door first. Tier 2: some data behind them. Tier 3: mixed or thin evidence. Tier 4: little or no proof. Worth doing? is our plain verdict. How sure is how confident we are in the rank. Rank weighs how much a tactic is likely to matter as well as how good the evidence is, so a Low can sit above a Medium. Rank 19 is a Yes that sits low only because measuring does not raise your odds.
| Rank | Tactic | Tier | Worth doing? | How sure | The evidence in a phrase |
|---|---|---|---|---|---|
| 1 | Let AI crawlers into your site | 1 | Yes | High | The platforms document it; the loss from blocking is unmeasured |
| 2 | Make pages readable without JavaScript | 1 | Yes | Medium | A 2024 log study, a platform note and a small test agree |
| 3 | Get into the search indexes AI tools use | 1 | Yes | Medium | Google documents it; others do not say which index they use |
| 4 | Lead with the answer and keep pages organized | 2 | Yes | Low | Lab tests on models, not pages; the platforms disagree on small answer blocks |
| 5 | Get your brand mentioned and linked | 2 | Yes | Medium | Strong patterns in the data, but no proof of cause |
| 6 | Keep content fresh, with honest dates | 2 | Yes | Medium | A lab test and two vendor studies favor recent dates; fake dates fooled ranking models |
| 7 | Publish original data and unique content | 2 | Yes | Low | Google asks for unique content; tests of AI-targeted rewrites are mixed |
| 8 | Earn press coverage, not just your own pages | 2 | Maybe | Medium | Studies split on who gets cited most |
| 9 | Add structured data (schema markup) | 3 | Maybe | Low | The platforms give mixed guidance |
| 10 | Keep pages fast and reliable | 3 | Maybe | Low | Correlation only |
| 11 | Publish honest comparison pages | 3 | Maybe | Low | Heavily cited by ChatGPT; one Google check found self-promotional lists often cited but not recommended |
| 12 | Take part in forums and Reddit | 3 | Maybe | Low | A top source for Perplexity and Google AI Overviews; OpenAI cited few Reddit pages, Anthropic none |
| 13 | Make your site easy for browser agents | 3 | Maybe | Low | Good practice, thin proof |
| 14 | Publish YouTube and video content | 3 | Maybe | Low | A strong pattern, but it measures mentions and citation shares vary widely |
| 15 | Let AI agents plug into your software (MCP) | 3 | Maybe | Low | Descriptions matter; listings unproven; software sellers only |
| 16 | Build reviews and directory listings | 3 | Maybe | Low | Mixed results |
| 17 | Translate pages and tag language versions | 4 | Maybe | Low | Language changes who gets cited; translation is untested |
| 18 | Get into AI training data | 4 | Skip | Low | No known way to raise it |
| 19 | Measure your AI visibility | 4 | Yes | Low | It does not raise your odds; single checks mislead |
| 20 | Serve Markdown versions of pages to agents | 4 | Skip | Low | Bot visits did not clearly rise in a randomized test |
| 21 | Get a Wikipedia article | 4 | Skip | Medium | Untested; Wikipedia discourages writing your own; earn independent coverage first |
| 22 | Add an llms.txt file | 4 | Skip | Medium | Several checks found no benefit |
Tier 1: Open the door first (ranks 1 to 3)
The platforms document these, or logs confirm them. We rank them first because the rest of this list matters much less if AI tools cannot reach, read and index your pages.
1. Let AI crawlers into your site
How sure we are of this spot: high.
AI assistants that search the web send out crawlers, programs that read your pages. OpenAI's search crawler is OAI-SearchBot, Anthropic's is Claude-SearchBot, and Perplexity says PerplexityBot should be allowed if you want to be included.
Anthropic says blocking its search crawler may reduce your visibility in user search results. This ranks first because nothing else works behind a closed door, not because anyone measured the gain. We found no study of what blocking costs.
Blocking is common. Cloudflare (vendor) reported that in 2025 AI crawlers were the bots most often blocked outright in robots.txt files, the files that tell bots which pages they may visit.
Do this: Ask your web developer to open your robots.txt file and allow the search crawlers. Check your firewall (the filter that blocks unwanted visitors) and your CDN (the service that speeds up your site) too, because either can block bots. Cloudflare, for example, lets you allow or block AI bots by category. The lines to add:
User-agent: OAI-SearchBot
Allow: /
User-agent: Claude-SearchBot
Allow: /
User-agent: PerplexityBot
Allow: /
If your file already has rules for all bots (they start with User-agent: *), ask your developer to copy any of them you still want into each block above. A bot with its own block usually ignores the all-bots rules.
OpenAI and Anthropic also run separate crawlers, GPTBot and ClaudeBot, that collect text to train their models. You can allow the search crawlers and block these.
2. Make pages readable without JavaScript
How sure we are of this spot: medium.
Many modern sites build their text in the browser with JavaScript, code that runs on the visitor's device. In December 2024, Vercel and MERJ (vendor) studied one month of Vercel's network logs, which record which bots visited and what they fetched. They found that OpenAI's and Anthropic's crawlers download JavaScript files but do not run them. Vercel sells hosting that sends finished pages from the server, so it has a stake here.
The study covered crawlers such as OpenAI's GPTBot and Anthropic's Claude crawler. For the search crawlers named above, treat it as a strong hint, not proof.
Gemini, which uses Google's crawler, and Apple's crawler do run JavaScript. A page built in the browser can look empty to a crawler that skips the script. Anthropic's documentation says its web fetch tool does not support JavaScript-rendered pages.
In a small test by SEO consultant Glenn Gabe, ChatGPT, Perplexity and Claude could not read a site that builds its pages in the browser, while pages built on the server worked.
Do this: Ask your developer to load your key pages with JavaScript turned off. If the main text disappears, have them send finished pages from the server instead (called server-side rendering). Google's crawler runs JavaScript, so this matters most for the other assistants.
3. Get into the search indexes AI tools use
How sure we are of this spot: medium.
Google says a page is eligible for its AI Overviews and AI Mode (its chat-style search) if it is indexed and allowed to show a snippet, with no extra requirements. Indexed means listed in Google's search database; a snippet is a short text preview.
What that means for ChatGPT and Claude is less clear. OpenAI says only that ChatGPT search sometimes partners with other search providers, and Anthropic's docs never name the search provider.
Ranking well in Google is not the whole story. In an August 2025 study of 15,000 narrow, specific questions, Ahrefs (vendor) found that on average only 12% of the pages ChatGPT, Gemini and Copilot cited also ranked in Google's top 10 for the same prompt. Perplexity overlapped more, at about 29%. Ahrefs attributes the gap to query fan-out, the extra searches an assistant runs behind the scenes.
So being indexed is the entry ticket, but no one can tell you exactly which index each assistant uses.
Do this: Check that your key pages are indexed by searching site:yourdomain.com in Google and Bing, or by opening Google Search Console and Bing Webmaster Tools, the free site reports each engine offers. Ask your developer to remove any noindex or nosnippet tag (code that hides a page or its text preview) unless you added it on purpose.
Tier 2: Some data behind them (ranks 4 to 8)
Studies find a link to being cited. Most are correlations, lab tests or vendor-run, and some point the other way, so treat the sizes as rough.
4. Lead with the answer and keep pages organized
How sure we are of this spot: low.
In a 2026 lab experiment (252,000 test runs, six AI models), the models most often cited first the text that closely matched the question or came earlier in the material they were given. A stated price and a recent date also helped.
A peer-reviewed paper first posted in 2023, "Lost in the Middle", found that the language models it tested answered most accurately when the relevant facts were at the start or the end of a long text. Neither study tests how a web page is laid out, so the link to page design is indirect.
Whether to break pages into small answer blocks is also unsettled. Microsoft recommends one- or two-sentence answers, lists, tables and question-and-answer blocks. Google says chunking, its word for splitting content into small pieces, is not needed for Google Search.
Do this: Put a direct answer at the top of each page, then the supporting facts, prices and lists. Write for readers first, with clear sections and lists. Do not create separate fragment pages just for bots.
5. Get your brand mentioned and linked
How sure we are of this spot: medium.
One of the strongest measured links to being named by AI is what other sites say about you. Ahrefs (vendor) studied 75,000 brands and found that how often a brand is mentioned across the web moves closely with how often AI mentions it. The correlation was 0.66 to 0.71, where 1 would be a perfect match.
SE Ranking (vendor) looked at ChatGPT citations and found the strongest patterns for referring domains (other sites that link to you), domain trust (a site authority measure) and traffic. Sites with 32,000 or more referring domains were 3.5 times more likely to be cited than sites with up to 200. The two studies measure different things, brand mentions in AI answers and page citations in ChatGPT, so they do not line up head to head, and neither shows cause.
A lab test with three AI models on simulated skincare products found that a known brand won every recommendation when specs were equal. It tests brand familiarity, not links or mentions, so it is only loosely related.
Do this: Ask partners, customers and industry groups to link to you and describe what you do, and earn mentions on credible outside sites in your field. Earn them naturally; do not buy, trade or fake them.
6. Keep content fresh, with honest dates
How sure we are of this spot: medium.
The same 2026 experiment found that a recent date helps content get cited. SE Ranking (vendor) found that pages updated within the past three months averaged 6.0 ChatGPT citations each, against 3.6 for other pages.
Ahrefs (vendor) found that URLs cited by AI assistants average 1,064 days old (about three years), against 1,432 days (nearly four years) for organic search results. So cited pages are fresher, but still years old, and ChatGPT leaned most toward recent pages while Google's AI Overviews cited older ones. The SE Ranking and Ahrefs figures are correlations; the experiment is a lab test.
Dates carry weight, which makes them easy to abuse. In a lab study posted as a preprint (a paper shared before expert review), seven AI models that re-sort search results by relevance ranked a passage higher when it carried a fake newer date. The text itself was unchanged. Larger models were less affected, but still moved.
Do this: Update key pages with real changes, show the true last-updated date and keep the last-modified date in your sitemap (the file that lists your pages for search engines) accurate. Bing says this helps it revisit changed pages. Never fake a date.
7. Publish original data and unique content
How sure we are of this spot: low.
Google lists unique content that only you can offer as a foundation for doing well in its AI features. The research on rewriting pages for AI is mixed. An early study, presented at a 2024 conference, found that rewrites adding sources, quotations and statistics raised how prominently the rewritten page appeared in AI answers by up to 40% in its tests.
Two later sets of tests, C-SEO Bench and SAGEO Arena, found that most tactics aimed at AI answers do not help and often hurt under realistic conditions. No study isolates original data on its own, so this ranks on Google's advice and a mixed research record.
Do this: Publish facts nobody else has, such as your own measurements, first-hand results and real customer outcomes, with clear sources. Do not add quotes or statistics to a page just to look citable.
8. Earn press coverage, not just your own pages
How sure we are of this spot: medium.
Rank 5 is about being mentioned across the web. This one asks which kind of source AI cites most: outside coverage you did not pay for, or your own pages. Studies split. A University of Toronto preprint found that AI engines cite outside coverage, such as news articles, far more than brand-owned or social content.
Yext (vendor) counted 6.8 million citations across ChatGPT, Gemini and Perplexity in four industries in mid-2025. It found that 86% came from brand-managed sources, meaning a brand's own websites and listings.
Profound (vendor) looked at 8,061 categories across eight assistants from April to July 2026. The typical category got 57% of its citations from brand sites, 30% from outside coverage and 15% from social sources (each figure is the middle value for its own source type, so they need not add to 100). Brand sites were 47% of ChatGPT's citations and 69% of Gemini's. The studies differ in assistants, industries and how they sort sources.
Do this: Pitch credible independent outlets when you have real news, but do not rely on them alone. Two vendor studies found that most AI citations go to brands' own pages and listings, so keep those accurate too. Neither study shows that more work there raises your citations.
Tier 3: Mixed or thin evidence (ranks 9 to 16)
Studies disagree, results flip by assistant, the audience is narrow, or the only support is a correlation. Test before you invest.
9. Add structured data (schema markup)
How sure we are of this spot: low.
Schema markup is code that labels the parts of a page, such as a product, a price or an FAQ, so software can read them. The platforms give mixed guidance.
Google says special schema is not needed for Google Search, though its 2025 list of ways to do well in its AI features includes structured data that matches the visible content. Microsoft recommends JSON-LD, one format for schema.
SE Ranking (vendor) found that FAQ schema did not help ChatGPT citations, in a correlation study rather than a controlled test. A slide deck from OtterlyAI (vendor) claimed large gains in Google's AI Overviews and AI Mode and drops on ChatGPT, Gemini and Copilot, with sample sizes mostly not shown.
Do this: Add accurate schema that matches what the page shows if your site platform makes it easy. It is aimed mainly at Google and Bing. Do not build extra markup just for AI, and do not expect FAQ schema to lift ChatGPT citations.
10. Keep pages fast and reliable
How sure we are of this spot: low.
SE Ranking (vendor) found that, among the pages it studied for ChatGPT citations, pages that showed their first content in under 0.4 seconds averaged 6.7 citations each, against 2.1 for the slowest pages. That is a correlation, not a test.
iPullRank (vendor) reports that a Profound analysis of 700,000 pages found that pages that failed or timed out on more than 75% of load attempts were cited about 18 times less often. Its reading is that such pages drop out of the running rather than being marked down. This is one firm's view of a sample from April 2026.
Google's 2025 list of ways to do well in its AI features includes good page experience, but gives no results.
Do this: Fix pages that error out or time out first. Speeding up pages that already work is optional.
11. Publish honest comparison pages
How sure we are of this spot: low.
Best-of lists are heavily cited. Ahrefs (vendor) found that best-of blog lists were 43.8% of the 26,283 pages ChatGPT cited across 750 prompts in three categories. Siege Media (vendor) found that business-to-business sites with more versus and comparison pages got more visits from AI search, a correlation of 0.65 across 116 sites, though brand strength probably explains part of that.
The risk sits with self-promotional lists. SEO consultant Lily Ray checked 100 business-software queries on three dates from April to June 2026. In the 80 that showed a Google AI Overview, she counted 323 citations of self-promotional lists. In about 69% of those 323 citations, the AI Overview cited the list but did not recommend the brand that wrote it.
Do this: Write honest versus pages for real alternatives. If you publish a best-of list, do not just rank yourself first. Google's spam policy now covers attempts to manipulate generative AI responses, so keep the list honest.
12. Take part in forums and Reddit
How sure we are of this spot: low.
Reddit's weight varies by assistant. In data from Profound (vendor) on 680 million citations from August 2024 to June 2025, Reddit was 46.7% of the citations going to Perplexity's ten most-cited sources and 21.0% for Google AI Overviews. Across all their citations, its share was much smaller: 6.6% for Perplexity and 2.2% for Google AI Overviews.
In a different sample, from March to July 2026, DEJAN (vendor) found that OpenAI's search pulled up 491,024 Reddit pages but cited only 3,012. Anthropic's search cited Reddit zero times in the sample.
Forums can also be abused. Researchers at Cornell Tech, in a preprint, showed that AI research tools can be steered through forum posts.
Adding as few as 13 words to one user-posted page, such as a Reddit thread, was enough. Across three open-source tools that write reports from web searches, those words were pulled into the research in 57 to 76% of runs and cited in 38 to 51% of reports. Treat that as a warning, not a tip.
Do this: Answer real questions in communities your buyers use, openly and within each forum's rules.
13. Make your site easy for browser agents
How sure we are of this spot: low.
Some AI products now browse the web and act for you, clicking and typing like a person. OpenAI says ChatGPT Atlas, its AI browser, uses ARIA roles and labels (tags that describe the parts of a page to software) to read page structure, and that accessibility improves agent performance. Google's web.dev says browser agents see sites through screenshots, raw HTML and the accessibility tree (a labeled outline of the page), and advises real buttons and links and stable layouts.
The one direct test we found is small: a July 2026 paper compared a single agent-ready design with a baseline built for humans, 150 runs each. On shopping tasks, an AI agent fully passed 134 of 150 tries on the agent-ready design and 74 of 150 on the baseline. That measures finishing tasks, not being recommended.
Do this: Use real buttons, links and form labels, keep layouts stable and avoid overlays that block clicks. It is good practice anyway.
14. Publish YouTube and video content
How sure we are of this spot: low.
Ahrefs (vendor) found that YouTube mentions of a brand track AI mentions of it more closely than anything else it tested, a correlation of about 0.737 across 75,000 brands. That measures mentions of a brand on YouTube, not the effect of publishing your own videos.
How often assistants cite YouTube depends on who is counting. Elmo (vendor) compiled six published figures for YouTube's share of AI citations that range from about 1% to more than 20%, and says the spread comes from different engines and different counting bases. In a 30 day sample of more than 100 million citations from OtterlyAI (vendor), Perplexity and Google AI Overviews accounted for most of the YouTube citations. Gemini and Copilot accounted for almost none of the YouTube citations (0.2% and 0.5%).
Do this: Make video if your customers want it anyway.
15. Let AI agents plug into your software (MCP)
How sure we are of this spot: low.
This one matters only if you sell software. MCP, the Model Context Protocol, is a way to let AI agents use your product's features, and an MCP server is the connector that offers them. OpenAI says the model picks which tool to call from the name, description and input format the server exposes.
In a lab experiment posted as a preprint, AI agents chose among servers that do the same job. Servers whose descriptions met the paper's own quality standard were picked 72% of the time, against a 20% baseline. We found no study showing that being listed raises recommendations. Anthropic ranks its directory by how many distinct Claude accounts sent a server any message in the last 30 days, so a listing alone does not earn placement.
Do this: If you already run or plan an MCP server, give each tool a plain name and a plain description of what it does. We found no study showing that building one raises how often you are recommended, so do not build one for that reason.
16. Build reviews and directory listings
How sure we are of this spot: low.
When ratings are supplied to the model, they sway its picks in two single-category lab tests. In a skincare test, a rival rated slightly higher took a known brand's lead. In an audit of twelve AI models choosing hotels, a top rating added 31.6 percentage points to a hotel's chance of being selected. Neither test checks whether earning reviews or listings on real sites changes what assistants recommend.
The software evidence is thinner and points the other way. One small study by DerivateX (vendor), run only on ChatGPT with search, found no citations of G2 or Capterra, two review sites for business software, among 169 cited URLs across 40 software categories. That conflicts with G2's own claims about its influence.
On searches for local businesses, Yext (vendor) found in late 2025 that Claude cites reviews and social sources several times more often than other models, with the gap varying by sector.
Do this: Keep your profiles and reviews accurate where your buyers look. Do not count on G2 or Capterra profiles alone to earn ChatGPT citations.
Tier 4: Little or no proof (ranks 17 to 22)
No independent test shows a lift, the tests that exist found nothing, there is little a site owner can do, or the tactic measures rather than improves.
17. Translate pages and tag language versions
How sure we are of this spot: low.
The language of a question changes which sites AI cites. Profound (vendor) studied 3.25 billion citations across 7 AI models and 14 countries and found that language shifts which domains are cited. That is a reason to think translation could matter, not evidence that it does.
The GEO Lab says no controlled experiment isolates hreflang, the tag that tells search engines which language version to show, and we found none for translation either. Bing says AI systems group near-duplicate pages and may show the wrong one. It lists hreflang and canonical links (tags that name the main version) among the ways to signal which page you want.
Do this: If you serve customers in other languages, translate for them and mark each language version with hreflang tags so search engines show the right one.
18. Get into AI training data
How sure we are of this spot: low.
AI models learn from huge collections of text gathered before release, called training data. In a test of 2,430 responses from Claude Code, an AI coding tool, across three Claude models, Amplifying (vendor) found that it most often wrote custom code. When it did name a tool, it sometimes picked one almost every time, such as GitHub Actions in 93.8% of its picks in that category, though picks varied by model.
Amplifying used Claude itself to decide which tool each answer named. The authors say they cannot tell whether the picks reflect a tool's quality or how often it appeared in the model's training data. So the study shows that models can favor some names heavily, but not how a site could earn that.
We found no study showing how a site can raise its place in training data. The only controls we found let a site opt out, using robots.txt rules aimed at AI training, such as the ones for OpenAI's GPTBot, Anthropic's ClaudeBot, Google-Extended, Applebot-Extended and CCBot. Google says blocking Google-Extended can also limit whether Gemini apps use your pages to answer questions.
Do this: Skip trying to raise it, because no method is known. Blocking training crawlers such as GPTBot is a separate choice: it tells those companies not to train on your pages, and no study we found shows it helps or hurts your citations.
19. Measure your AI visibility
How sure we are of this spot: low.
Measuring does not raise your odds by itself, which is why it ranks low. It does show whether your visibility is rising or falling.
A single check misleads. SparkToro and Gumshoe.ai (vendors) ran the same 12 prompts in 2,961 runs with 600 volunteers, and identical prompts returned different brand lists nearly every time. Averaging over many prompts helps. OtterlyAI (vendor) found that a brand's coverage score, the share of tracked questions that mention it, swung about 73% less when based on 100 questions than on 10, in one 30 day dataset.
Both big search engines now report on AI activity. Google Search Console, Google's site report, launched AI Overviews and AI Mode reports on June 3, 2026 and rolled them out to every site by August 31, 2026. They count how often your pages appeared, not clicks or queries.
Bing Webmaster Tools added an AI Performance report in public preview in February 2026. It shows how often your pages were cited and what searches the AI ran. It covers only Copilot, Bing AI summaries and select partners, and Bing says it does not show ranking or a page's role in an answer.
Do this: Open the AI reports in Google Search Console and Bing Webmaster Tools, and track many prompts over time. Do not trust a single run of one prompt.
20. Serve Markdown versions of pages to agents
How sure we are of this spot: low.
Some agent tools ask for Markdown, a plain-text format, and it can take far fewer tokens, the small text units AI tools are limited by. Cloudflare (vendor) showed one page dropping from 16,180 tokens to 3,150 with its own Markdown feature, and Claude Code's fetch tool prefers Markdown. We found no study showing a payoff in citations.
Profound (vendor) ran a randomized test that served Markdown to AI crawlers on 381 pages across six sites over about three weeks. Bot visits rose about 16% on average, but the test could not rule out chance, and the typical page gained just one visit. The test measured crawling, not citations.
Google's Danny Sullivan said on the Search Off the Record podcast, as Search Engine Roundtable reported on January 9, 2026, that Google Search does not want people making separate versions of their content for AI tools.
Do this: Skip separate Markdown pages made only for bots. If you run developer documentation, serving Markdown on request can save agent tokens, but no citation gain is proven.
21. Get a Wikipedia article
How sure we are of this spot: medium.
Profound (vendor) found that Wikipedia was 7.8% of all ChatGPT citations in its data, which covers 680 million citations from August 2024 to June 2025. Semrush (vendor) measured a different thing, the share of tested ChatGPT responses that cited Wikipedia, and reports that it fell from roughly 55% to under 20% in September 2025. Treat the size as unsettled.
We found no controlled study that tests whether having an article raises your mentions. One vendor talk, from OtterlyAI, reported citation gains from Wikipedia with sample sizes mostly not shown. Wikipedia's own notability rules say a company earns an article only through significant, independent, reliable coverage by outlets with no tie to it, and paid media and press releases do not count.
We rank this a skip because an article follows independent coverage and cannot be ordered up.
Do this: Do not write or edit your own article directly. Earn coverage from outlets with no tie to you first, and let others write about you. If your company already has an article and you are connected to it, say so on the article's talk page (its discussion page) rather than editing it yourself.
22. Add an llms.txt file
How sure we are of this spot: medium.
An llms.txt file is a plain-text guide to your site written for AI tools. SE Ranking (vendor) compared about 300,000 domains and found no link between having the file and being cited, without weighing the quality of the files. Ahrefs (vendor) studied server logs for 137,210 domains in May 2026 and found that 97% of the roughly 38,000 valid llms.txt files got no requests.
EZY (vendor), whose platform generates the file, logged 12 weeks of fetches on 83 sites. OpenAI's crawler requested robots.txt 3,990 times and llms.txt 7 times, and only Meta's crawler fetched llms.txt often. Google says llms.txt is not needed for Google Search.
The strongest claim in favor comes from Mintlify (vendor), which hosts documentation sites. It says llms.txt cut AI agent errors by about 90% on the documentation it hosts, based on its own data.
Google's Chrome team has an experimental llms.txt check in Lighthouse, its site audit tool. It calls the file an emerging convention and says agents may spend more time crawling a site without it.
Do this: Do not expect citations from it. Skip it unless you run developer documentation that coding agents read; the only support for it there is a documentation host's own figure and a Chrome audit note. If you already have one, it is fine to leave.
What we would do first
- Ask your developer to check robots.txt and your firewall settings, and let the search crawlers in.
- Turn off JavaScript and confirm your key pages still show their text.
- Confirm those pages are indexed in Google and Bing, and open the AI reports in Search Console and Bing Webmaster Tools (rank 19).
- Put the answer first on each key page, add facts only you have and keep your dates honest.
- Earn mentions on credible outside sites (rank 5). The evidence is correlation, so treat it as a bet, not a guarantee.
- Skip llms.txt, bot-only Markdown pages and writing your own Wikipedia article.
Hand this list to whoever runs your website and start with step 1 this week. The order will change as new tests come in. This one reflects what was published through September 2026.