A buyer asks ChatGPT or Claude to find a physio nearby who treats running injuries, and the assistant goes off and does the looking. Nobody reads the results page except the agent. This lesson follows what it does at each step, using the few measurements that exist, and ends with the order to fix things in.
The agent answers from memory first
An agent doesn't always search. The agency Synscribe recorded three research runs of one AI agent, Claude Code, in mid-2026 and published what the traces showed in its report How AI Agents Actually Search. In all three runs the first answer came from what the model already knew, with no web search at all, and the live search started only after the user pushed back. That's three runs on one agent, a small pilot, so read it as a direction rather than a rate.
For a business it means two things decide your chances. What other sites already say about you, in directories, reviews and articles, shapes the answer from memory before your own site is opened. And when the agent does go and look, your pages have to hold up. The lesson on brand mentions and entity SEO covers the first part; the rest of this one covers the second.
Most results are never opened
In the busiest of those runs, Synscribe counted 215 different websites in the agent's search results and about 13 that it opened, roughly 6%. That comes from a single run, so treat it as the shape of the problem, not a benchmark.
The agent decides what to open from the search result itself: the title, the description, and whether the page looks like it answers the question. A title that names the service and the place earns the open, and a slogan rarely does. URLs, titles and meta descriptions shows how to write one.
Then the page has to load for the agent. AI search agents name themselves when they visit: OAI-SearchBot and ChatGPT-User for OpenAI, Claude-SearchBot and Claude-User for Anthropic, PerplexityBot and Perplexity-User for Perplexity. A robots.txt rule or a firewall that turns them away takes you off the list without telling you. Many agents also read the HTML as it arrives rather than running the page's scripts, so text that appears only after scripts run may never be seen.
It drops claims nobody else backs
In the same run, the agent checked the claims it had collected against other sources and threw out about 48% of them, Synscribe reports. Again, that's one run on one agent.
The claims at risk are the ones on your homepage: years in business, jobs completed, a star rating, an award, a licence, a guarantee. A claim only your own site makes is the easiest to drop. A claim that a review platform, a licensing register or a news article also states is one the agent can confirm.
So list your five boldest claims and find where else each one is stated. Where nothing backs a claim, put it somewhere it can be checked: ask for reviews on the platform your buyers use, link the register entry for your licence, or soften the claim to what you can show.
It needs a price, a way in and a name
Once you're on the shortlist, the agent compares and acts. It looks for what you charge or how to get a quote, a phone number, email or booking link that works, and who runs the business. A template phone number left in the footer, a contact form with unlabelled fields, or prices that exist only inside an image can stop it at the last step.
One more thing can make an agent distrust a site outright: text hidden where people don't see it, in alt text, labels or invisible elements, that gives the AI instructions. Microsoft's security researchers call this AI recommendation poisoning. However it got onto your pages, remove it before anything else.
What to fix first
Work through the steps in the order the agent meets them:
- Let the agents in. Check robots.txt, your firewall or bot protection, and that the homepage carries no noindex tag. These usually take minutes.
- Put the facts in the page. Prices or how to get a quote, real contact details, and the people behind the business, written as text in the HTML.
- Make your claims checkable, and remove any hidden instructions.
- Then the extras. An llms.txt file and markdown copies of key pages help agents that look for them. Google's own guide to AI features says you don't need new machine-readable files to appear in Search, so treat these as plumbing for agents rather than a ranking lever. The free llms.txt generator writes the file and checks a live one.
What to take away
- An agent often answers from memory first, so what other sites say about you counts before your own pages are read.
- Being in the results is not being opened: in Synscribe's one measured run the agent opened about 6% of the sites it found, so titles, descriptions and access decide whether you're read.
- A claim only your site makes is the easiest one to drop, so back your boldest claims somewhere an agent can check.
- Fix in the agent's order: let it in, put prices and contact details in the page, make claims checkable, then add llms.txt.
Next
For how the same assistants decide which passages to quote in an answer, read Writing for AI citations.