How AI engines decide which companies to name

Two ways an engine knows about you

Memory. Large language models carry a compressed picture of the web from their training data. If you have a long, consistent public record, an engine may name you without searching at all. That picture changes slowly, and the engine rarely says where it came from.

Search. For anything current or specific, the engine searches, reads pages and writes an answer, often with links. Google describes the technique behind its AI features as query fan-out: one question becomes many searches across related subtopics, and the answer is assembled from what comes back. So there is no single keyword to win. The engine reads the whole neighbourhood of the buyer's question.

Which of the two an engine relies on, and when, explains most of what we saw in our study.

Three engines, three salespeople

In our study of 240 AI conversations about money, the three engines behaved like three different salespeople.

• ChatGPT cites. It searched in every first and second answer and linked a source in 96% of all answers, mostly public bodies. It kept company names back until the buyer asked, and put a firm forward early in only 7% of conversations.
• Gemini asserts. It linked sources in 63% of answers, named firms and prices from memory, and never once flagged them as unverified. It put a firm forward before being asked in 66% of conversations.
• Claude hedges. It linked sources in 81% of answers and, in 43 of 56 conversations, said it had not checked the firms it named. In 11 it declined to name advisers at all.

What the engine can read decides what it can quote. When the buyer asked about cost, ChatGPT reached 27 firms through a price page and Claude 25. Gemini, which searched in only 15 of 56 of those answers, reached 11.

One answer is a sample

Answers are generated fresh each time, and the pages an engine reads can differ between runs. Ask the same question twice and you can get a different answer.

For one expat couple with identical details, the maximum mortgage each engine gave moved across ten runs:

Lists of companies move the same way. Across the three engines, 252 brands were named, 27 of them on all three, and about 40% appeared once and never again. That matches SparkToro's research earlier this year, which found less than a 1 in 100 chance of the same list of brands appearing twice.

So a single screenshot of an AI answer tells you little. Your position is the share of answers you appear in, per engine.

Four things that change underneath you

The machinery moves on its own clock, and a business can lose ground without changing a page.

• Models. What the systems can do. Google moved AI Mode to a new default model worldwide in May 2026, and a new model can describe the same business differently.
• Retrieval. What gets read before an answer is written, and from which sources.
• Infrastructure. What you need to have set up to be reachable. From 15 September 2026, Cloudflare applies new default settings for AI bots on new sites and unchanged free plans.
• Policy. What platforms allow. Google's spam policies now name attempts to manipulate generative AI responses as spam.

Each change calls for one of three responses: act now, keep an eye on it, or safely ignore it.

The lens applied: what you can do yourself

Each row is a finding from the study, read through the technology, with a check or fix you can make yourself. The last column names the gate and state in our framework: Found (does AI name you?), Understood (does it get you right?) and Actionable (where does it send a buyer who is ready?). The overview explains each state.

Try it yourself: a 20-minute technical check

1. Open your robots.txt and your CDN or hosting settings. Check whether AI crawlers are allowed on the pages you want read.
2. Pick the five facts a buyer needs most: price, coverage, certifications, delivery or availability, contact. Check each one is in plain text on a page, not only in a PDF or image.
3. Ask one buyer question five times on each engine. Write down every name and every link.
4. Count how often you appear per engine. That share is your starting point.
5. Note which sites each engine links to for your field. Those are the pages that shape your answer.
6. Put a reminder in your calendar to repeat it after the next major model or policy change.

Why one lens is not enough

The technology tells you what an engine can read and how it behaves. It does not tell you which questions your buyers ask, or whether what the engine says about you is right. Read the other two lenses: how buyers decide inside an AI conversation and the version of your company AI tells. For all three applied to one industry, see One study, three lenses.

Frequently asked questions

How does ChatGPT decide which companies to recommend? In our study it searched for most answers, linked public sources heavily and held company names back until the buyer asked who to go to. When it did name firms, it often reached them through a page that answered the buyer's question directly, such as a price page.

Why does AI give different answers to the same question? Because answers are generated fresh each time and the sources an engine reads can differ between runs. That is why one answer is a sample, and why a position should be measured as a share across many runs.

Should I block or allow AI crawlers? It is a trade-off. Blocking them keeps your content out of AI answers built from live search. Allowing them lets engines read and cite you. Decide per type of page, deliberately, rather than leaving a default in place.

Is this the same as SEO? Some foundations carry over: a readable, well-structured site helps both. The difference is that an AI engine writes one answer from many sources, often third-party ones, instead of ranking your page in a list.

Where do you stand with each engine?

If you would like to know what ChatGPT, Gemini and Claude say when your buyers ask, we are happy to talk it through. In a 20-minute call we look at how a potential customer would describe the problem you solve, what each engine says when they do, and whether optimising for the AI buyer journey is worth your attention right now. Sometimes the answer is that it can wait, and we will say so.

Book a 20-minute call

Sources

1. CheckMarque AI, "Three AI engines, eight Dutch money questions," 10 October 2026.
2. Google Search Central, "AI features and your website." https://developers.google.com/search/docs/appearance/ai-features
3. SparkToro, "AIs are highly inconsistent when recommending brands or products," 28 January 2026. https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility/
4. Google, "Google Search's I/O 2026 updates," 19 May 2026. https://blog.google/products-and-platforms/products/search/search-io-2026/
5. Cloudflare, "Cloudflare Allows the Agentic Internet to Flourish with a Simple Philosophy: Your Content, Your Rules," 2026. https://www.cloudflare.com/press/press-releases/2026/cloudflare-allows-the-agentic-internet-to-flourish-with-a-simple-philosophy-your-content-your-rules/
6. Google Search Central, "Spam policies for Google web search" (updated 28 August 2026). https://developers.google.com/search/docs/essentials/spam-policies