AI assistant recommendations: why it names your competitor and not you

You and your competitor probably both qualify for AI assistant recommendations

Eligibility, usually. The vendors document the same conditions for both of you. Google says a page must be indexed and eligible to be shown in Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, that there are no additional technical requirements, and that meeting them doesn't oblige Google to crawl, index or serve the page. It adds that AI Overviews appear only where its systems judge them additive, so they often don't trigger. OpenAI says sites opted out of OAI-SearchBot won't be shown in ChatGPT search answers, though they can still appear as navigational links. Perplexity says PerplexityBot is designed to surface and link websites in its results. Anthropic says disabling Claude-SearchBot may reduce a site's visibility and accuracy in search results.

If you both clear those conditions, the documents have nothing more to say. None of these vendors explains how one eligible page is preferred to another, and we won't pretend to fill the gap. We can ask a smaller question: what could an assistant check, by reading, about each of you, and where would you differ?

What can an assistant check about a firm by reading?

As we read the documents, an assistant can check the passage, the firm's identity, and whether anyone else agrees. Start with the passage. Google's guide to snippets says they are created from page content and chosen to preview the part that best relates to a user's specific search, so different searches produce different snippets. OpenAI's search tool lists opening a page and finding text inside it as separate actions. Both describe a pipeline that works at the level of passages. A page helps if it holds a stretch of text that answers the question as asked, and being about the right subject isn't enough.

Identity comes next. Google says adding Organization structured data to your home page can help it disambiguate your organization from others, and that a sameAs property can point to your profile pages on other sites. Its AI features page says no special markup is needed, so read this as the vendor describing a problem, not prescribing a fix. A system that reads pages from across the web must decide that the name on one is the firm on another. One name and one sentence of description everywhere makes that easy. Variants make it uncertain.

Corroboration comes last. Google's page on helpful, reliable, people-first content asks whether a page shows clear sourcing and background about its author or site, and whether someone researching the site would come away thinking it well trusted or widely recognized as an authority on its topic. It describes Google's ranking systems, not the choice of names in an answer. It does show that agreement from outside the firm is the kind of evidence the vendor treats as relevant to a source.

These checks are our reading, not a documented formula. No vendor says they decide the outcome, or how they would be weighed.

Where eligible firms part company

The table below sets the checks side by side for a pair of firms. We built them for illustration; they aren't records of any real business. Both pass the documented conditions, and every difference between them is something a system could establish by reading. Read down the rows, mark where the firms part, and then hold each row up against what your own pages publish today.

How can one equally able firm win the recommendation?

Take a retail business asking which firms can connect its store stock system to its online shop. The first firm has a long record of this work. Its page is titled with a service category, opens on a view of where commerce is heading, and describes its method in the vocabulary of its brand. It uses a short name on its site and its full legal name elsewhere. The only outside descriptions of it are directory entries, each written once and never revisited.

The second firm is no better at the work. Its page is titled with the problem, opens with a paragraph saying what it connects to what and for whom, and lays out the order the work is done in. A trade body's member page and a published talk describe it in the same terms. An assistant composing an answer has a sentence to lift from the second firm and can check that other sources agree. For the first it has nothing to lift and nothing to confirm.

Both firms are composites. This is how we'd expect it to go, an inference from how firms describe themselves and not an observation of any assistant. It suggests the firm that gets named is often the one that was easier to describe, and that isn't always the one that's better at the work.

Which of these can you change this quarter?

They move at different speeds. The passage is the quickest because it's your own text: you can rewrite a page in a day and test it the next. Identity takes longer, because it lives in every place the firm describes itself: the site, social profiles, speaker biographies, directory entries, press boilerplate, and the legal name on documents that end up online. One person should hold the approved name and a single approved description, the way a style guide holds a brand, and every outside party that asks for a description should get that text.

Corroboration is the slowest, and the only one you can't do alone. A trade body, a publication or a client describes a firm when it has a reason to, and the reason is usually work it can point to. The straightforward route is to do work worth describing and make it easy to describe correctly. A firm that tries to manufacture the mentions is working on the wrong problem, and nothing we read gives a basis for thinking it would help.

In what order do you look for the gap?

The diagram below gives the order we use. It starts with access, because that part is documented and certain to fix: a blocked crawler, or a page not eligible to be shown, ends the inquiry. Next comes the opening of the page, which is cheap to change and easy to test. Only then does it turn to the harder job of making the firm describable the same way across its own pages and other people's.

Run each test on more than one day and with more than one assistant before you conclude anything. Google notes that its AI Overviews and AI Mode can draw on different models, so the links each shows can differ, and it's sensible to assume the same between vendors.

What if the recommendation never came from your pages?

Then none of the above applies, and nothing the vendors document says how often it happens. Anthropic's documentation says Claude answers directly, without searching, when a request draws on stable knowledge, and OpenAI's guide says the model chooses whether to search. An answer given without a search names firms from the model's own recall. OpenAI also states that each crawler setting is independent, so a site can allow its search crawler and disallow its training crawler. The pages we read don't say whether training access changes what a model recalls about a firm, and we don't claim it either way.

So don't conclude too much from one answer, and don't copy the competitor's page. Firms that share a description are harder for a reader to tell apart, not easier.