How AI assistants choose which brands to recommend: the five gates

An assistant does not rank brands. It passes a question through a sequence of gates, and a brand that fails any one of them is absent from the answer no matter how well it would have done at the next. The gates, in order, are: does the answer call for a name at all, can the system reach and retrieve something about you, does more than one independent source say the same thing, is your claim specific enough to repeat, and then the tie break that decides the order. This page walks through each gate, what fails at it, and what fixes it.

Gate 1: does the question call for a name

Many category questions come back as guidance with no companies in them. Whether names appear is decided mostly by how the question is phrased, and that is useful to know before you conclude anything about your own visibility.

• A question with a constraint in it, a budget, a team size, a location, a compliance requirement, usually produces names.
• A broad definitional question usually produces an explanation.
• Comparison and shortlist phrasing almost always produces names, and often three of them.

So the first thing to establish is whether the question you are testing is a naming question at all. A prompt that never names anybody cannot measure whether you are named.

Gate 2: reach and retrieval

Here the two kinds of answer separate. A fetched answer searches the live web while you wait. A training answer is written from what the model already holds. Ask the assistant whether it searched and which pages it read, and you will know which one you are looking at.

For a fetched answer, four things decide whether you are retrievable, and all four are cheap.

• Named crawler access. A robots file written for Googlebot years ago says nothing about the AI agents. OpenAI publishes its tokens at platform.openai.com/docs/bots and states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers.
• Text in the HTML. Words that appear only after JavaScript runs are words the fetcher may never see.
• No wall in front of the page. Logins, interstitials and aggressive bot management at the CDN all look the same to a fetcher.
• One answer per URL. Retrieval works on passages. A page that answers one question offers a clean passage. A page covering nine offers nine diluted ones.

For Google's AI surfaces the requirement is published: Google's documentation on AI features states that a page must be indexed and eligible to be shown in Search with a snippet to appear as a supporting link, with no additional technical requirements.

Gate 3: corroboration

This is the gate most well built websites fail, and it explains more absences than any other.

A system assembling a recommendation is pulling from several sources at once. Pew Research Center's analysis of Google AI summaries found that 88 per cent of the summaries examined cited three or more sources, and that Wikipedia, YouTube and Reddit were the most frequently cited domains in both summaries and standard results. A claim that exists only on your own domain has one source behind it. The same claim on a review platform, a trade directory and a trade publication has four, and four sources that agree is what a recommendation is made of.

That is why the work splits into two halves that feel unrelated. Writing pages makes you retrievable. Being described by other people makes you recommendable. The evidence on the second is in why listings and reviews move AI answers.

Gate 4: specificity

A model will commit to repeating a fact it can stand on. It will not commit to a superlative.

What your page says	What an assistant can do with it
Industry leading solutions for modern business	Nothing. There is no claim to repeat
Insolvency software for resolution professionals, used on 400 cases, priced per case	Repeat the category, the scale and the pricing model
Trusted by businesses nationwide	Nothing
Offices in Leeds and Manchester, covering 60 postcode districts, same day callout before 2pm	Answer a location question and a response time question

Counts, prices with the month you checked them, dates, named accreditations with issuing bodies, and a category description in your buyers' words. Those are the sentences that get quoted, and most marketing copy contains none of them.

Gate 5: the tie break and the order

Once several businesses pass the first four gates, three things tend to decide who is named first.

• Fit to the constraint. The company whose public text matches the specific constraint in the question, rather than the biggest company in the category.
• Weight of agreement. How many independent sources describe you in the terms the question used.
• Freshness. A page with a recent date on a question about current prices or current law beats an undated one.

Size helps at this gate because large companies are described in more places, which is a corroboration advantage rather than a preference for brands. We tested the pattern in do AI answers favour big brands, and the practical consequence is that a specific, well corroborated niche claim beats a vague large one.

Why it names your competitor

Work the gates in order and the cause is nearly always one of three, in this frequency.

• Nothing outside your own domain describes what you do. Gate three. The most common and the slowest to fix, which is why it should start today.
• No single page of yours answers that question. Gate two. A week of writing.
• The fetcher cannot read your site. Gate two again, and the cheapest of all to fix.

Check them in reverse cost order: access first, then pages, then corroboration. The diagnostic is in why your competitor is recommended by AI and you are not, and the per engine behaviour in how ChatGPT picks which companies to name.

This page is the mechanism: the gates a brand passes through before it is named. For the method that moves each gate, see LLM optimization. For the sources an answer is built from, see what AI reads when it answers a buying question.

How AI assistants choose which brands: asked and answered

Is there a ranking formula I can optimise against?

There is no published ranking for an AI answer, and the gates above are observable and testable, which makes them a practical substitute. Run your own prompts, record who is named and what is cited, and the pattern in your category becomes visible within one round.

Can I pay to be named?

Advertising placements exist and are labelled as such, separately from the generated answer. Nobody sells a place inside the organic answer, so treat any such offer as a reason to ask for the contract in writing.

Does structured data make an assistant pick me?

It helps a system understand what your entity is, and Google asks that your markup match the visible text. The plain words on the page are what gets quoted, so treat markup as supporting work rather than the main lever.

Why do two runs of the same question name different companies?

Answers are generated at the moment of asking, so variation between runs is normal. Measure across several runs and report a range, and treat a change as real only when the ranges stop overlapping.

How long does it take to get named?

Access fixes show up in fetched answers within days of the next crawl. New pages that answer a buyer question tend to appear over a few weeks. Corroboration from third party sources builds over months, which is why it is the half to start first.

What to do first

• Pick three buyer questions with a constraint in them and ask each one three times in two assistants, naming nobody.
• Ask each assistant whether it searched and which pages it read, and write those URLs down.
• Check robots.txt and your access log for the named AI crawlers, and fix anything blocked by accident.
• Count the sources outside your own domain that describe what you do. If that count is low, that is gate three and it is your priority.
• Rewrite one page so the first two sentences answer the question and contain a count, a price or a date.