What changed
Three years ago a buyer typed a phrase into Google, saw ten links and chose. Now the same buyer asks ChatGPT, Perplexity, Gemini or Google's AI Mode a full question, reads a short answer with a few names and a few cited sources, and chooses from those. The names in the answer are the shortlist. A firm that is not named was not rejected; it was never seen.
The mechanism behind that answer is different from classic ranking, and in 2025 and 2026 enough data was published to describe it. This article sets out what the industry studies of millions of AI citations and Google's own documentation agree on, what they contradict, and what a small firm with a modest site should do in what order.
What the engines draw from
Third-party sources, mostly. When an assistant recommends a vendor it cites the vendor's own website only a minority of the time. Across several 2026 datasets covering ChatGPT, Perplexity, Google AI Overviews and AI Mode, the most cited sources for "best firm for X" questions were directory list pages, review platforms, comparison articles, forums and journalism. One study of supplier-selection prompts found sector directories supplied over forty percent of citations, with the directory's list page cited far more often than any single firm's profile on it. Another, across twenty-five million links, found earned media of all kinds made up the large majority of citations and paid content a fraction of a percent.
Mentions, not links. A 2025 study of tens of thousands of brands found the number of times a brand was named on other websites correlated with its AI Overview visibility about three times more strongly than its backlink count did. The top quartile of brands by mentions received many times the citations of the next quartile, and a quarter of brands had none at all.
Pages that answer first. Analysis of over a million ChatGPT answers found that most citations came from the first third of a page, and that focused pages were cited more than long guides. Pages written in definitive language, dense with named things, and containing the questions they answer did better.
Recency. Content under a few months old is cited more often, and Gemini weights recency strongly because it grounds its answers on live Google search.
Sub-questions. Google's Search Central documentation explains that AI Overviews and AI Mode may use query fan-out: issuing several related searches across sub-topics and sources, then assembling an answer from the pages that cover each. A 2026 analysis of millions of AI Overview citations found only around a third of cited pages ranked in the top ten for the original query. The others ranked for the sub-questions.
A consistent entity. Engines reconcile what they read about a business against what they have read elsewhere. The same legal name, description, founder, founding year, address and phone on the site, in its Organization markup and on every directory profile is what lets them do that. Buyer research finds that a mismatch between a firm's website and what is said about it elsewhere reads as risk.
What does not help
AI text files. The much-promoted llms.txt file is not needed, according to Google's documentation, which says in plain words that no new machine-readable files, AI text files or markup are required for its AI features. A 2026 study of over a hundred thousand sites found that almost all such files received no requests at all, and that the handful of fetches that did occur came mostly from developer tools, not search engines. It does no harm. It does nothing.
Schema as a citation lever. Structured data has real uses: Organization markup feeds brand panels, Breadcrumb and Article markup are supported for ordinary search features. But a 2026 controlled study of nearly two thousand pages that added structured data, against a control group, found no increase in citations on ChatGPT or AI Mode and none on AI Overviews. Google's own guidance is consistent with that. FAQ rich results were removed for all sites in May 2026 and HowTo rich results in 2023; the markup still parses and earns nothing. Keep the question-and-answer content. Drop the expectation that the markup does the work.
Bought mentions. Since May 2026 Google's spam policies explicitly cover attempts to manipulate generative AI responses, and treat paid, manufactured or coordinated brand mentions the way bought links have long been treated. Combined with the finding that paid content is a fraction of a percent of citations, the case for buying your way in is closed: it does not work, and it can now cost you.
Chasing head terms. Nearly all queries that trigger an AI Overview are informational and long. "Software development company" rarely triggers one; "how much does a Magento to Shopify migration cost" does. A twenty-five page site cannot win the head term and does not need to.
Pages in bulk. City-swap pages, keyword variants and generated content are covered by the scaled content and doorway policies, which since 2026 also apply inside AI features.
A practical order of work for a small firm
- Open the door. Check robots.txt and, more importantly, the CDN or firewall. The search-index crawlers of ChatGPT, Perplexity and Claude are separate from the training crawlers, and a "block AI bots" switch usually blocks both. Blocking training is a choice; blocking the search crawlers removes you from citations. Verify Search Console and Bing Webmaster Tools and submit the sitemap to both.
- Restructure the pages you have. One page, one question. The question as the heading; the direct answer (what, for whom, what it costs, how long) in the first screen in plain text; the detail below. Prices and facts as text, never in images or calculators. Merge pages that could swap headings without anyone noticing.
- Write the missing questions. List the sub-questions your buyers ask: cost, duration, comparison, what breaks, what is included. Each gets one focused page from your own facts. These are the fan-out pages.
- Make the entity consistent. One name, one description, one founder, one founding year, one address and phone. Put it on the about page, in Organization markup, on your Business Profile, and on every directory profile. Fix every place it differs.
- Earn presence where the engines read. Find the two or three directories cited for your category and complete your profile with verified reviews, aiming for the list page. Find the comparison articles cited for your questions and learn their inclusion criteria. Take part honestly in the forum threads where firms like yours are discussed. Ask every finished client for a review on the platform that counts.
- Publish one real number. An original dataset, even a modest one, stated in a sentence with its method, is the cheapest citable thing a small firm can produce. Then tell the people who write the comparison articles about it.
- Measure monthly. Same questions, same engines, same conditions, recorded.
Steps one to four are engineering and editing, and a small firm can do them in weeks. Steps five and six are slower and compound over months. Nothing on the list requires a special file, a plug-in or a payment to anyone.
How to measure yourself for free
Write down the five questions your buyers would ask an assistant, phrased the way they would type them. Open a fresh, logged-out session of ChatGPT with search on, Perplexity, Gemini, and Google in a region your buyers are in. Ask each question in each engine. For every answer, record three things: were you named, was your site cited, and if not, which sources were. Save the answers with the date.
You now have a twenty-cell baseline. Repeat it monthly under the same conditions. If you would rather we did the first run for you, the free AI visibility check does exactly this across four engines from a US location and returns a written report in two working days. The paid work that follows, if you want it, is described on the AI search visibility page.
A note on scepticism: AI answers vary from run to run, and most of the studies cited here measured large brands and software categories. The direction of the findings is consistent across independent datasets; the exact percentages are not yours. Measure your own questions and trust your own table.
Questions people ask
Does ranking first on Google mean we will be cited in AI Overviews?
It helps and it is not enough. Industry analysis of millions of AI Overview citations in 2026 found that only around a third of cited pages ranked in the top ten for the typed query. The rest ranked lower or ranked for a related sub-question that the engine generated itself. Google's documentation describes this as query fan-out. A page that answers one narrow question plainly can be cited for a broad query it would never rank for on its own.
Is it true that an llms.txt file helps AI engines find us?
No. Google's documentation says no AI text files are needed for its features, a Google representative has said publicly that no AI system uses the file, and a 2026 server-log study across more than a hundred thousand sites found that almost all such files were never requested by anyone. It does no harm, and it does nothing. The hour is better spent completing a directory profile.
Should we buy a placement in a 'top agencies' list?
No. Industry studies put paid and advertorial content at a fraction of a percent of AI citations, so it rarely works. And since May 2026 Google's spam policies explicitly cover attempts to manipulate AI responses, treating paid or planted brand mentions the way bought links are treated. Earned placement on a list that selects on real criteria is a different thing, and it is worth asking what those criteria are.
How many reviews do we need, and does the rating matter?
Count matters more than rating. In a 2026 study of ChatGPT recommendations in business software categories, the number of reviews on a review platform correlated with being recommended, while the average rating showed almost no correlation. The practical reading: a steady flow of genuine reviews on the platforms the engines read for your category does more than a perfect score on a handful. Ask every finished client, in writing, and make it easy.
How often should we re-measure?
Monthly is enough to see movement without reacting to noise. AI answers vary from run to run, so a single check proves little; a monthly run of the same questions under the same conditions shows the trend. Record the prompts, engines, location and date each time. If you change pages or complete profiles, expect on-site changes to show within four to ten weeks and third-party presence to build over three to six months.