How AI Models Choose Which Brands to Recommend: The Signals That Matter in 2026

Which makes the useful question a different one. Not how do we rank our brand in ChatGPT, but what would make our company a credible, well-supported answer to the things our buyers actually type.

Nobody outside these companies knows the exact weighting. OpenAI says ChatGPT Search uses several factors meant to surface relevant, reliable information. Google says its AI features lean on the same core Search ranking and quality systems it already runs. That's less than a formula, but it's enough to work with.

This piece walks through what the public guidance supports, what it doesn't, and what's worth doing either way. If you'd rather not run the testing by hand, Xtrusio does the same pass across ChatGPT, Gemini, Claude and AI Overviews and keeps each cycle comparable to the last.

The question comes before the brand

Nothing is picking winners from a table. It's trying to answer the thing somebody just typed, using whatever it can find that fits.

Take three questions that look almost the same:

• What are the best project management tools?
• What are the best project management tools for construction companies?
• What is the best project management software for a 20-person construction firm that needs mobile time tracking?

They're related. They do not have the same right answer. A globally known platform may fit the broad first question but be a weaker fit for the very specific third one. That difference is easy to miss if you only measure broad category or branded questions.

So visibility has to be looked at per question and per use case. Category-level data can be too broad. Brand-name searches answer a different question because the person already knows who you are.

Here's what that looks like in practice. In one Xtrusio audit, a scoring tool rated a company 33 out of 100 for AI visibility. That score was built on brand-name searches: prompts where the company had already been named. We then ran 25 questions a real buyer in that category would type, across three AI assistants, which produced 75 answers between them. The company appeared in none of them. Competitors appeared in most.

Nothing about the company changed between the two tests. Only the questions did. One audit isn't a study, and a bigger sample might soften the split. The point holds either way. A brand-name score tells you how you're described once somebody knows your name. It says nothing about whether you get named when they don't.

The prompt is not always the search

When ChatGPT uses web search, OpenAI says it may rewrite what somebody typed into one or more tighter searches, and run further ones after seeing what comes back. Google describes something similar in its AI features, where a single question can fan out into several related searches.

Say a buyer asks which contract management platform suits a large healthcare group. Answering that properly means touching contract software, healthcare use cases, enterprise rollout, security, compliance, integrations and what customers say once they're live. Having clear, credible information across those areas gives a search or AI system more relevant material to work with than simply repeating ‘contract management platform’ on your homepage.

Which is why five questions are worth asking about every buyer question that matters to you:

1. Is the brand genuinely relevant to this use case?
2. Is there content proving it, rather than asserting it?
3. Can anyone outside the company confirm it?
4. Is the information current?
5. Can a crawler actually reach it?

That gets you further than adding the same target keyword to one more blog post.

An AI has to work out what you sell

A system can't confidently connect you to a recommendation if it can't work out what you do. Plenty of sites make that harder than it needs to be.

A homepage says it's transforming tomorrow through intelligent digital experiences. It reads well. It's also unclear what the company sells, who buys it and what problem goes away afterwards.

Compare it with this: Acme builds inventory forecasting software for multi-location retailers, connecting sales, stock and supplier data so retail teams can forecast demand and cut stockouts. Same length. This one carries a category, an audience, a capability and a business problem, and every one of those is checkable.

Consistency is not the same as identical wording

Your positioning turns up in more places than your homepage. Product pages, the About page, LinkedIn, software directories, partner sites, review platforms, press coverage, interviews, conference bios.

None of that needs to be word-for-word identical. It does need to stop contradicting itself. Say your site calls you an AI visibility platform. One directory files you under SEO agencies. Another has you down as social listening. An old interview describes you as content automation. Ask what category this company belongs to and the honest answer is that nobody is sure.

Pick the associations you want to own

Write down the things you want your company tied to, then go and check whether the web agrees. For an accounts payable company that list might read: mid-market finance teams, multi-entity invoice processing, NetSuite and SAP integrations, less manual invoice handling.

The job isn't repeating a keyword thirty times. It's making the link between your company and a real buyer need easy to follow.

The source has to fit the claim

There is no public trust score. Different AI products run different retrieval systems and pull from different places.

What the major platforms do say out loud is that they're after relevance, reliability and content quality. OpenAI says ChatGPT Search ranks results on several factors meant to help people find relevant, reliable information, and that it doesn't publish a weighting or promise placement. Google says its AI features pull current, relevant pages through its core Search systems before generating an answer. Its wider Search guidance points at experience, expertise, authoritativeness and trust. Trust is the one it calls most important.

So the question worth asking isn't which domain authority number ChatGPT prefers. It's why this source would be a credible place to back up this particular claim.

Third-party validation, without the volume game

Your site should explain what you sell. Buyers rarely stop there, and neither does anything trying to answer on their behalf. Industry publications, analyst firms, customer reviews, credible directories, comparison sites, conference talks. None of it sits under your control. That is exactly what makes it worth something.

The temptation is to treat that as a numbers exercise. Google's guidance for its AI features warns against exactly that. Manufacturing mentions is unlikely to pay off the way marketers hope, because the same quality and spam systems are still running underneath. A smaller number of credible, relevant references can be more useful than a large batch of low-quality placements on unrelated sites.

Original evidence gives people something to cite

A lot of marketing content repeats the same definitions, benefits and advice.

That may still be useful, but it gives another writer or buyer less reason to reference one particular page. Original research, testing and real customer evidence give them something more specific to work with.

Google's guidance asks for unique, non-commodity content built on first-hand experience rather than a rework of what's already published. The same principle is useful beyond Google: if you want people or systems to reference your work, give them something specific worth referencing.

The kinds of thing that qualify:

• proprietary research and annual benchmarks;
• anonymised platform data;
• customer surveys;
• experiments and hands-on product testing;
• pricing studies and market indexes;
• implementation timings from real rollouts;
• case studies with numbers in them.

Say you analyse a few thousand AI answers across twenty software categories and publish how often the cited sources are vendor-owned versus independent. That's a new fact in the world. Another writer can lean on it. A journalist can quote it. A customer can use it in a shortlist meeting. Compare that with the eleventh version of seven reasons AI search matters for business.

Show how you got the number

Don't just publish a surprising figure. Say what was measured, how big the sample was, when it was collected, which platforms were tested, what you assumed and where the method falls down. Google's guidance asks creators to be clear about who made something, how, and why. Readers want the same thing, and the ones who check are usually the ones worth convincing.

If you want a worked example of turning this into a repeatable exercise against named rivals, the competitor benchmarking playbook covers the mechanics.

None of it counts if nothing can reach you

Before doing anything advanced, make sure the important pages can actually be reached.

Website settings or security tools can accidentally block search and AI crawlers even when the content itself is fine. This is usually quick to check and easy to overlook.

For ChatGPT Search, one specific check is whether OAI-SearchBot is allowed. Your technical team should also confirm that the website's security tools are not rejecting the crawler.

You do not need to audit every page first. Start with the pages that matter most to the buyer questions you want to appear for: product pages, use-case pages, comparisons, research and important guides.

Ask the technical team to check a few things. Whether important pages are blocked. Whether the wrong version of a page is being treated as the main one. Whether pages load correctly. Whether links and redirects work. And whether anything important sits behind a login.

There is no file that gets you in >Google does not require a special AI file or special AI schema for its AI search features. New technical experiments can still be tested, but they should not distract from the basics: reachable pages, clear information, useful content and claims that can be checked.

Recommendations are scenarios, not standings

Appearing today doesn't mean you hold a position. Change the user, the location, the wording or the requirement and the answer can move.

ChatGPT Search can use general or precise location where geography matters, and OpenAI says query rewriting can fold in relevant context to build tighter searches. So treating AI visibility like one fixed results page will mislead you. Think in scenarios instead.

Build a fixed set of buyer questions covering the main stages of the buying journey. A few dozen well-chosen questions can tell you far more than one or two screenshots.

Illustrative structure. Build the set around how your buyers actually choose, then leave it alone so the next run is comparable.

Then record the same things every run. Whether you appear. Which rivals do. Where you sit in the answer. How your product gets described, and whether that description is right. Which sources got cited and what domains they sit on. Then watch how all of it moves. A first pass through what appearing in ChatGPT actually involves is worth having before you build the set.

A citation count is not a ranking >Microsoft's AI Performance report in Bing Webmaster Tools shows citation activity and the queries behind it. Microsoft says plainly that a citation count doesn't represent ranking, authority or position inside an answer. Worth holding onto for any AI visibility programme. One response is only one example. What matters is the pattern across repeated runs.

What we can honestly say about the signals

There is no published weighted formula for ChatGPT brand recommendations. Anyone presenting a precise formula is adding assumptions that the platforms have not published.

From OpenAI's own Search guidance we can say two things with a straight face. Relevance and reliability matter in search-grounded results. And the system may rewrite a request into several searches before it answers. Past that, the honest answer is that assumptions are assumptions.

The parts you can improve are much simpler:

• a clear explanation of what your company does and where it fits;
• specific product information;
• content for real buyer situations and use cases;
• first-hand evidence and original research;
• credible coverage on websites you do not own;
• accurate and consistent company information;
• important pages that search and AI crawlers can reach;
• customer proof with real numbers.

Every one of those helps a human evaluate you too. That's a decent test for whether an AI optimisation tactic is worth the money. If it only works on machines, be suspicious of it.

And if you don't yet know why AI systems keep naming your competitors, start by finding out rather than publishing more. Compare the questions where each brand appears, open the sources behind those answers, and work out which evidence gaps you can realistically close this quarter. That beats another AI SEO checklist. You end up with a clearer view of why one brand may be easier to support for a particular question, and a short list of what to improve next.

See which questions name you and which do not

Some AI visibility tools rely heavily on searches for your own brand name. Xtrusio starts from the category questions where nobody has mentioned you yet, then reports who came back, which sources were cited in the answer and how the picture shifts between runs. Engines covered: ChatGPT, Gemini, Claude and AI Overviews. The waitlist is open while the product is being built.

Join the Xtrusio waitlist

Sources: OpenAI, Overview of OpenAI crawlers, for OAI-SearchBot and the published crawler ranges. OpenAI's ChatGPT Search documentation, for ranking factors, query rewriting, location handling and the absence of guaranteed placement. Google Search Central, AI features and your website, for query fan-out, technical eligibility, the advice against AI-specific files and the warning on manufactured mentions, and Creating helpful, reliable, people-first content, for original non-commodity content and the who, how and why guidance. Microsoft Bing Webmaster Tools, for the AI Performance report and its note that citation activity does not represent ranking, authority or placement. The audit figures are from one Xtrusio client engagement and are a single observation, not a study. Product behaviour and documentation change, so check the live sources rather than this page.