Why an AI doesn’t recommend you

The shift, in three figures

It has happened faster than most people planning websites have registered.

• 51% of business software buyers now begin their research inside a chatbot rather than a search engine. In April 2025 it was 29%.
• In a survey of nearly 18,000 business buyers, generative AI was named the most meaningful research source by twice as many people as any other, ahead of vendor websites.
• Traffic arriving from AI converts at 14.2% against 2.8% from organic search. Roughly five times better, for the obvious reason: it arrives pre-qualified.

The consequence is the part worth sitting with. The first impression of your company now forms inside an answer you cannot see, before anybody opens your site. By the time they arrive, they have a shortlist, and either you are on it or the visit is a formality.

They do not cite your website, and they do not agree with each other

This is where the intuition fails. The instinct is: write the site well, the AI reads it, the AI recommends you. For questions about who to hire, that is not how it works.

For those questions the engines cite third parties almost exclusively, and they pick different ones:

• Google AI Overviews takes 97% of what it cites from the top twenty organic results. Which means classic search work still decides this one.
• ChatGPT is roughly 48% Wikipedia, then Reddit, Forbes, review sites. It leans on training data first and only searches when it judges that it needs to.
• Perplexity leans on Reddit, YouTube, analyst firms. Freshness matters enormously: content under thirty days old receives about 3.2 times more citations.

And the number that stops any single strategy from working: only about 11% of the domains cited by ChatGPT are also cited by Perplexity. These are not one channel with four front doors. They are four different systems that happen to look the same.

The thing being sold that does nothing

If somebody has offered to sell you an `llms.txt` file, here is what is published about it.

Google updated its documentation in June 2026 to say the file has no effect on Search or on AI Overviews, and that Search ignores it. The reasoning given is hard to argue with: a manifest written by the site itself cannot distinguish between sites, because every site would claim to be the best one.

The measurements agree. Ahrefs tracked 137,210 domains carrying the file and 97% received no requests for it at all in a single month. SE Ranking ran the file against AI citation data across roughly 300,000 domains and found no correlation.

Some assistants do read it, and it costs one file, so there is no reason not to have one. There is every reason not to pay for it as a strategy. We publish one on this site and we do not count it as anything.

What actually moves it

If the engines cite third parties, then the work is getting third parties to describe you. Not more pages on your own domain.

The clearest published number in this whole field comes from Clutch, whose data shows providers with fifty or more reviews receiving around 63 times more AI citations than providers with none. Going from zero reviews to one nearly triples them. A verified profile averages 8.6 times more than a free one.

Read the first and last of those together and you get the practical shape: the enormous multiplier needs years, and the first meaningful jump is from nothing to one. That first one costs a phone call to a client who already likes you.

The same logic runs through the rest. Getting included in somebody else's comparison article. Being visible where your industry actually discusses suppliers. Having a presence on the platforms these systems draw from. None of it is on your website, and all of it is what a system reads when it decides whether you exist.

What your website still has to do

It has a job. It is a narrower one than it used to be.

It answers the problem question. Somebody who does not yet know that companies like yours exist is searching for their problem in their own words, and a page that answers it properly is still how they find you. That is what a blog is for, and it feeds Google AI Overviews too, since 97% of what that cites comes from ordinary search results.

It answers the verification question. Somebody who has your name from an AI answer or a recommendation comes to check. A site with no external trace, no profiles, nothing corroborating it, reads exactly as it is: a page making claims about itself.

What it does not do any more is win the shortlist. That is decided elsewhere, and pretending otherwise is how companies spend a year rewriting a homepage that was never the problem.

How to check where you stand, this afternoon

No tool required and it takes twenty minutes.

Open ChatGPT, Perplexity, Gemini and Claude. Ask each the question a buyer would actually ask, in their words, not yours. Not "what is a good supplier of integrated logistics solutions" but "who do I call to move pallets between Spain and Portugal".

Note who comes up. Then ask where that came from, because most will tell you. Those sources are your target list, and they will not be the same list on each engine.

Then ask each one about your company by name. That answer is what a buyer sees after somebody recommends you, and it is often the more revealing of the two.

Repeat it every couple of months. It is the only measurement here that is free, and it is the one that matters.

What we would not spend money on yet

This field has grown an industry faster than it has grown evidence, and some of what is being sold is priced as certainty.

Tools that promise to make you appear in AI answers are mostly selling monitoring, which is genuinely useful, wrapped in a claim about influence, which is not established. Monitoring is a report. It is not a lever.

Anyone promising a specific position in a specific engine is selling something nobody can deliver, since the systems change without notice and disagree with each other by design.

What is evidenced is prosaic and cheaper: third-party reviews, being present where your buyers already talk, and content that answers real questions well enough that somebody else cites it. That has been the answer for a while. The engines changed and the answer did not.