Schema, llms.txt and AI Search in 2026: What Actually Gets Businesses Cited

Schema markup and llms.txt files are not what gets a business cited by ChatGPT or Google's AI answers. Controlled 2026 studies found schema had no measurable lift and almost no llms.txt file was ever read. What works: plain answers near the top of each section, real numbers with sources, server-rendered pages and a strong reputation across the web.

Key takeaways

• Adding schema did not move ChatGPT or Google AI Mode citations in a matched study of 1,885 pages.
• 97% of llms.txt files received zero requests in a month, and Google says AI text files are not needed.
• Clear answers, cited statistics, quotations, crawlable HTML, reviews and third-party mentions are what the evidence supports.

Every few months a new AI search fix gets packaged and sold to business owners. In 2025 and 2026 the two big ones were schema markup and a file called llms.txt. Both sound technical and both are easy to bill for. This guide looks at what independent data says about each one, and what to spend your time on instead.

Why are agencies selling schema and llms.txt as AI search fixes?

Because they are easy to install, easy to show on a report and sound like the machine-readable answer to a machine problem. The logic feels right: AI reads code, so give it more code. The problem is that the controlled tests run in 2026 do not show either one changing who gets cited.

Schema markup is structured data, usually JSON-LD, that labels facts on a page: this is a business, this is its address, this is a product price. llms.txt is a proposed plain-text file at the root of a website that summarizes the site for large language models.

Neither idea is harmful. The issue is being told they are the reason a business will or will not show up in ChatGPT. That claim can now be checked against data.

Does schema markup get a business cited by ChatGPT or Google AI?

Not measurably. An Ahrefs study compared 1,885 pages that added JSON-LD schema with matched pages that never did. Citations rose about 2.4% in Google AI Mode and 2.2% in ChatGPT, results the authors could not separate from zero. In AI Overviews, cited pages fell 4.6%.

The method matters here. Ahrefs used a difference-in-differences test, which compares how cited pages changed in the 30 days before and after schema was added against similar pages that did nothing. The schema additions happened between August 2025 and March 2026.

The AI Overviews drop was statistically significant, but the authors are careful about it. Both groups were already declining before the change, so they do not claim schema caused the fall. The fair reading is simple: schema did not help, and it may not have hurt either.

There is one caveat worth knowing. Every page in the study already had more than 100 AI Overview citations, so the results describe pages AI systems already knew about. Even so, the same article notes that 53% of AI-cited pages have schema. That is correlation, and well-built sites tend to have both schema and good content.

Google says the same thing in plain words. Its guidance on AI features states there is no special schema.org structured data you need to add to appear in AI Overviews or AI Mode.

Does an llms.txt file help AI tools find a business?

Almost never, based on the server logs. Ahrefs looked at 137,210 domains in May 2026. About 28% had an llms.txt file, and 97% of those files got zero requests that month. Of the requests that did happen, AI bots made only 19.5%.

The rest came from SEO audit tools, general web crawlers, tech profiling tools and other bots. GPTBot, OpenAI's training crawler, made about 4.5% of all requests. AI retrieval bots, the ones that fetch pages to answer a live question, made just 1.1%.

Ahrefs also points out that a fetch does not prove a file was read or used. So even these small numbers are a ceiling on real use, not a floor.

Google's AI features documentation is direct on this one too: you do not need to create new machine-readable files, AI text files or markup to appear in its AI features. If someone is charging a monthly fee to maintain an llms.txt file, that money can go somewhere more useful.

Is ranking in Google's top 10 still enough to get cited?

No longer on its own. A March 2026 Ahrefs analysis of 863,000 keyword results and about 4 million AI Overview URLs found that only 38% of cited pages also ranked in the organic top 10. In July 2025 the same measure was about 76%.

The likely reason is query fan-out. Google's AI splits one search into many related sub-questions, runs searches for each one, then builds an answer from the pages that show up across those searches. A page that answers several of those sub-questions well can be cited even when it does not rank for the main keyword.

In the same data, 31% of citations ranked between positions 11 and 100, and 31% sat outside the top 100 entirely. YouTube made up 5.6% of all cited URLs.

For a business, this is good news. You do not need to beat the biggest site on one hard keyword. You need clear, specific pages that answer the smaller questions buyers actually ask: cost, timing, process, comparisons and problems.

Myth vs evidence: what does the data actually say?

Most of the popular AI search advice fails when tested, while a few plain writing and reputation habits hold up. The table below puts each common claim next to the strongest evidence available in 2026, with the source for each so you can check the numbers yourself.

What actually makes AI tools cite a page?

Content that is easy to quote and easy to trust. The GEO research paper from Princeton and IIT Delhi, presented at KDD 2024, tested nine writing changes. Adding quotations lifted visibility about 41%, adding statistics about 31% and citing sources about 28%. Keyword stuffing made pages slightly worse.

The researchers measured how much of a page's text showed up in AI-generated answers, adjusted for where it appeared. The biggest gains went to pages that were not already on top. When a page ranked fifth in search results, citing sources raised its visibility by 115.1%, while the top-ranked page lost about 30% on average.

That lines up with how these systems work. An AI answer is built from short passages it can lift and attribute. A passage that states a clear fact, gives a number and names where the number came from is simply easier to use.

In practice, that means a few habits on every page:

• Answer first. Put a direct 40 to 60 word answer right under each heading, then add detail.
• Use question-shaped headings. Match the way buyers phrase their questions in a chat box.
• Add real numbers. Prices, timelines, counts and percentages, each with a named source.
• Quote people who know. An owner, a technician or a published expert, with a name attached.
• Keep it current. Put a visible date on the page and update facts when they change.

Why does server-rendered content matter for AI crawlers?

Because most AI crawlers only read the HTML your server sends first. Vercel found that OpenAI's GPTBot, OAI-SearchBot and ChatGPT-User, Anthropic's ClaudeBot and PerplexityBot do not execute JavaScript. Anything your site loads by script after the page opens may be invisible to them.

Google's Gemini is the exception, because it uses Googlebot's rendering. Apple's crawler also renders pages. But a page that only works for one crawler is leaving the others out.

Common problems include reviews loaded by a widget, prices pulled in by a script, tabs and accordions filled by JavaScript, and schema injected by a tag manager. A quick test is to view the page source in your browser. If your key facts are not in that raw HTML, many AI tools cannot see them.

This is one reason we build client websites in Webflow, which serves pages as finished HTML. Whatever platform you use, the rule is the same: important text belongs in the page itself.

How do reviews and third-party mentions affect AI recommendations?

They matter more each year, because buyers now ask AI tools who to hire. BrightLocal's Local Consumer Review Survey 2026 of 1,002 US adults found 45% now use AI tools for local business recommendations, up from 6% a year earlier. That makes AI the third most-used source, behind Google and Facebook.

Recency counts as well. In the same survey, 74% of consumers look for reviews written in the last three months, and 32% look for reviews from the last two weeks, up from 20% the year before. A steady flow of fresh reviews tells both people and machines the business is active.

An AI answer about who to hire is assembled from many sources, not only your own site. Review platforms, directories, news stories, industry associations and forum threads all feed into what a model believes about a business. If your name, address, phone, hours and services differ across those places, the model has less reason to trust any of them.

So the work outside your website matters: keep business facts identical everywhere, ask every happy customer for a review, and earn mentions on sites your buyers already read.

Should a business still add schema markup in 2026?

Yes, for Google search features, not for AI citations. Article, LocalBusiness, Product, Review and similar types can still earn rich results such as star ratings, prices and business details in Google. Those improve click-through. Just do not buy schema as an AI visibility service.

Some schema types are losing value. Google stopped showing FAQ rich results in Search on May 7, 2026, and is removing the related Search Console reports. The markup can stay on your pages, but it no longer earns extra space in results.

Two rules keep schema useful. First, it must match the visible text on the page, which Google's own guidance asks for. Second, it should be in the HTML the server sends, not added later by a script, so every crawler can read it.

What should a business do first to show up in AI answers?

Start with the pages that already earn money and the questions your buyers ask before they call. Fix how those pages answer, make sure every crawler can read them, then work on reviews and mentions. That order puts effort where the evidence is strongest and costs the least.

1. List 20 to 30 buyer questions. Pull them from sales calls, emails and the People Also Ask box.
2. Test them in AI tools. Ask ChatGPT, Gemini and Google AI Mode, and record which businesses get named.
3. Rewrite your top pages. Use question headings, answer-first paragraphs, numbers and named sources.
4. Check the raw HTML. Make sure prices, services, reviews and contact details are in the page source.
5. Clean up your listings. Use the same name, address, phone and hours on every profile.
6. Build a review habit. Ask every satisfied customer, and reply to every review.
7. Earn outside mentions. Local press, trade associations, supplier pages and podcasts all count.
8. Repeat the test monthly. Track whether your name shows up more often over time.

Frequently asked questions

Does schema markup help with ChatGPT citations?

Not in any measurable way so far. Ahrefs tracked 1,885 pages that added JSON-LD schema against matched control pages and found ChatGPT citations rose about 2.2%, a change that could not be told apart from zero. Schema is still worth having for Google rich results, but it is not an AI citation lever.

Should I delete my llms.txt file?

There is no need to delete it. It costs almost nothing to keep. Just do not pay for it or expect results from it. Ahrefs found 97% of llms.txt files got zero requests in a month, and Google says you do not need AI text files to appear in its AI features.

Is SEO still worth doing if AI answers the question?

Yes. AI tools still pull from search indexes, and Google's own guidance says normal SEO practices are what make a page eligible for AI Overviews and AI Mode. What changed is the target: you now need to rank for the many sub-questions behind a query, not only the main keyword.

How fast can a business improve its AI visibility?

Page fixes such as answer-first paragraphs, cited statistics and server-rendered text can be shipped in a few weeks. Reviews, mentions on other sites and consistent listings build over months. Measure by asking the same set of buyer questions in ChatGPT, Gemini and Google AI Mode each month and recording who gets named.

Is FAQ schema still worth adding?

Not for Google's search results. Google stopped showing FAQ rich results on May 7, 2026. The markup can stay on your pages without harm, but the visible questions and answers are what matter, because both people and AI tools read the text on the page, not the hidden code.

The bottom line

AI search rewards the same things good salespeople do: clear answers, real proof and a good name around town. Schema and llms.txt are fine to have, but the data says they will not get you cited. Spend the budget on content, crawlable pages and reputation instead.

If you want a second opinion on where your business stands, Book a free call with Mike, or run the free 60-second check to see how AI tools see your site today.

Sources

• Ahrefs: Does adding JSON-LD schema increase AI citations? (May 2026)
• Ahrefs: We analyzed 137K sites: 97% of llms.txt files never get read (June 2026)
• Ahrefs: 38% of AI Overview citations pull from the top 10 (March 2026)
• Aggarwal et al., GEO: Generative Engine Optimization, KDD 2024
• Vercel: The rise of the AI crawler (December 2024)
• Search Engine Journal: Google drops FAQ rich results from Search (May 2026)
• BrightLocal: Local Consumer Review Survey 2026
• Google Search Central: AI features and your website

Keep reading

• How to win AI search with answer engine optimization
• Invisible to the machine: AI search visibility in 2026
• Building brand fame beyond SEO in the AI era