Why AI Assistants Recommend Your Competitor, and How to Change It

There is a specific kind of bad afternoon that is becoming common. Somebody on the leadership team asks an assistant to recommend a vendor in your category, watches it name three companies, and none of them is you. Then they forward the screenshot.

The instinct in the room is always the same: the model is wrong about our product. It is not, because it never had an opinion about your product in the first place. It read what other people published and summarised the consensus it found. Your product was not judged and found wanting. It was not in the pile.

That distinction matters, because it points at completely different work.

How a recommendation is actually assembled

When an assistant is asked who the best providers in a category are, it does something closer to a literature review than an evaluation.

It draws on what it absorbed in training, which is broad, unattributable and slow to change. For many questions it also retrieves live pages and writes from those. Then it produces a shortlist that reflects, roughly, which names appear most consistently in the material it considers credible on that topic.

A model cannot try your product. It can only repeat what has been written about it. A worse product written about often, recently and clearly will beat a better one that nobody has described.

That is an uncomfortable sentence for anyone who has spent three years building something genuinely superior. It is also the whole opportunity, because published consensus is a thing communications teams have always been able to move.

The five places recommendations come from

In practice, shortlists are built from a small number of source types, and four of the five sit on domains you do not control.

Roundup and best-of articles. “Best X tools for Y”, published by trade titles, review sites and content marketers. These are disproportionately influential because their structure is unambiguous: a list of names, with reasons attached, in a format that is trivially liftable. If your category has ten of these and you appear in two, you have found your problem.

Review platforms. Structured ratings, categories and user reviews. The structure is the point: a platform that says what a product is, who uses it and how it scores gives a model something concrete to repeat.

Comparison and alternatives pages. “X vs Y”, “alternatives to Z”. Often published by competitors, which means your category’s comparison landscape may currently be written entirely by people with a reason to frame you unfavourably.

Community threads. Forums and discussion sites where people ask each other what to buy. These carry real weight, partly because they read as genuine experience and partly because some engines have direct access arrangements to that content. You cannot fabricate your way into these credibly, and attempting it is the fastest way to become a cautionary thread of your own.

Your own site. Real, but weaker than the rest on any contested claim, because it is the one source with an obvious incentive.

Notice that four of the five are earned. This is why AI recommendation work belongs next to PR and community rather than inside a website project.

Why incumbents win by default

Three compounding advantages, none of which are about product quality.

They have been written about for longer, so they occupy more of the training data. They are the default comparison point, so even a comparison article arguing against them mentions them twice as often as the challenger. And they appear in older roundups that keep getting cited long after they stopped being accurate.

The practical consequence: a challenger’s first goal is not to win the recommendation, it is to be in the consideration set at all. Being named as one of four is an enormous improvement over being absent, and it is a far more achievable target than displacing the category leader in the first sentence.

Diagnosing your own case

Before doing anything, find out which of two problems you have, because they need different work and different timelines.

Run the category prompt several times in fresh temporary chats, following the method in how to check what ChatGPT says about your brand. Then look at the pattern.

If the same competitor is named in every single run, that name is sitting in the model’s memory. Changing it is slow work measured against model releases, and the route is sustained publication volume rather than any single placement.

If the names shuffle between runs, the answer is being assembled from retrieval, and the sources are findable. Ask the assistant which sources support the recommendation. When it browsed, it will tell you. That list is your brief, and this version of the problem can genuinely move within weeks.

If you are named but qualified, look at the adjectives. “A good option for smaller teams” is a positioning problem that got written down somewhere, and it is usually traceable to one or two influential articles that framed you that way early.

What actually moves a recommendation

In rough order of effect per unit of effort.

Get into the roundups that already rank. Find the best-of articles that come up for your category, check which you are missing from, and approach the publishers the way you would for any other earned placement. Many of these are updated periodically, and being absent is frequently just an oversight rather than a judgement.

Fix your presence on review platforms. Complete the profile, get the category right, and encourage real customers to review honestly. The category field matters more than people expect: being listed under the wrong one makes you invisible to the question your buyers are actually asking.

Publish the comparison your competitors wrote about you. If the only “you vs them” page in existence is on their domain, a model has exactly one framing available. An honest comparison, including where the other option is genuinely the better fit, is more citable than a one-sided one, because it reads as reference material rather than as sales copy.

Be described the same way everywhere. One consistent sentence about what you do, used in your boilerplate, your profiles and every briefing. Scattered inconsistent descriptions are how a model fails to connect the references to you at all.

Show up where buyers ask each other. Honestly, under your own name, answering questions rather than pitching. Slow, unglamorous, and it compounds.

What does not work

Asking the model to remember you. There is nothing to correct in a live chat that persists for anyone else.

Buying your way into an organic recommendation. It is not a placement that is for sale, and anyone promising it is selling something else.

Astroturfing the community threads. It is detectable, it is increasingly detected, and the downside is a permanent public record of you doing it.

Mass-producing thin comparison pages. Fifty auto-generated “X vs competitor” pages give a model nothing worth quoting and mark your domain as low-substance on exactly the topic where you needed credibility.

Measuring the change

Two numbers, tracked against a fixed prompt set and a recorded date.

Presence rate: in what share of category prompts are you named at all? This is the one that should move first, and it is the one that matters most commercially.

Recommendation rate: when the assistant names a single best option, how often is it you, and who wins when it is not?

Run each prompt several times and report the spread, not the best run. A recommendation that appears in one run out of five is a real and fragile result, and reporting it as a win will cost you credibility the first time somebody in the room tries it themselves.

The wider frame for all of this, including what to do when the problem is accuracy rather than absence, is in AI reputation management. Watching it across engines and over time is what AI reputation tracking does.

Common questions

Why does ChatGPT recommend a competitor instead of us?

Almost always because the competitor appears in more of the sources a model leans on when it builds a shortlist: roundup articles, review platforms, comparison pages and community threads. The model is summarising a consensus it found rather than making a judgement about product quality, so the fix is to change what that consensus looks like.

Does being better than a competitor make an AI recommend you?

No, and this is the hardest part to accept. A model cannot evaluate your product. It can only repeat what published sources say, so a worse product that is written about more often, more recently and more clearly will be recommended ahead of a better one that is not.

What kind of pages influence AI recommendations most?

Third-party roundups and best-of lists, review platforms with structured ratings, head-to-head comparison pages, and community threads where people answer each other's buying questions. All four are on other people's domains, which is why this work sits closer to PR and community than to website optimisation.

How long does it take to change what an AI recommends?

Where the answer is assembled from live retrieval, changes can surface within weeks of new sources appearing. Where it comes from the model's training memory, expect a much slower clock measured in model releases. Running the same prompt several times shows you which of the two you are dealing with.

Can you pay to be recommended by ChatGPT or Perplexity?

Not in the way you can buy a search ad, and anyone offering guaranteed placement in an organic AI recommendation is describing something that does not exist. What you can legitimately do is earn presence in the sources those engines read, which is ordinary publishing and communications work.