Why AI recommends your competitor instead of you

A client watched an assistant answer a simple question about their category and name three competitors. Not them. The company asking had the better product, the longer track record, and the happier customers. None of that appeared in the answer.

The instinct in that moment is to treat it as a verdict. It is not. An assistant recommends the company it can most confidently connect to a category, and confidence is built from evidence the system can read, not from how good you actually are. A competitor gets named instead of you when the model has stronger evidence for them in one of four places: they are retrieved when the question is asked, their identity is clear and consistent, their pages state the answer directly, and other sources describe them the way you want to be described. Each of those is a gap you can find and close. This is an evidence problem wearing the costume of a quality problem.

Recommendation is retrieval, not judgement

Start with what the system is actually doing, because the fix follows from it.

When a browsing assistant answers "who are the best providers of X", it does not open a scoreboard of quality. It reformulates the question into one or more searches, retrieves a set of candidate pages, reads them, and composes an answer that names the companies it can most confidently associate with the category in a form it can use. When an assistant answers from training data alone, without browsing, it surfaces the entities most strongly associated with that category across everything it was trained on. Neither process asks which company is best. Both ask which company is most confidently and most legibly connected to the topic.

That distinction is the whole article. Being better than a competitor does not put you in the answer. Being more legible to the system does. A company can be the strongest operator in its market and still be invisible to the model, because the qualities that make it strong, the delivery, the relationships, the results, are not things a crawler reads. What a crawler reads is your pages, your structured data, your presence in a search index, and the way other sources on the web describe you.

The reason I find this framing useful with clients is that it turns a demoralising moment into a diagnosable one. "The AI likes my competitor more" is not actionable. "The model has stronger evidence for my competitor in these three specific places" is. Legibility is buildable, and it is measurable, which means the gap between you and the recommended competitor is a list of things to fix rather than a judgement to accept.

There are four places the evidence gap opens. They sit in sequence, and each is a prerequisite for the next. Work them in order.

Reason one: you never entered the candidate set

You cannot be recommended if you were never retrieved. This is the first filter, and it is the one most often missed, because it fails silently and it has nothing to do with the quality of your argument.

For any answer built by browsing, the model assembles a set of candidate pages, usually from a conventional search index, and works only from that set. If your page is not in the set for the query the assistant ran, you are not in the running, regardless of how good the page is. The competitor is simply present in the index for that query and you are not.

The causes here are technical and unglamorous. Content that is not in the server-rendered HTML, so a crawler that does not execute your JavaScript sees an empty page. Directives in your robots file that exclude AI crawlers, sometimes added in 2023 or 2024 on the reasonable-at-the-time view that your content should not train models, and never revisited now that the same lines also keep you out of the answers that would cite you. Slow or unreliable responses that cause a fetch to time out. A thin conventional search presence for the terms your buyers actually use, which starves the candidate set before the model even reads anything.

The signature of a retrieval failure is specific: the assistant describes your category correctly, names competitors fluently, and never mentions you, even when you prompt it narrowly toward your own niche. If you are absent even when the question is practically about you, the problem is almost always that you were never retrieved. This is usually the cheapest of the four to fix and the most commonly overlooked. I have written separately about the crawler-access half of this in is your site blocking the AI crawlers that could cite you, and it is worth checking your own directives before you assume the problem is anywhere more sophisticated.

Reason two: the model is more confident about who your competitor is

Suppose you clear retrieval. The model can find you. It still has to know who you are with enough confidence to stake a recommendation on you, and this is where a great many companies quietly lose to a smaller competitor.

These systems build an internal representation of entities: who you are, what category you belong to, what you do, who you serve. The confidence of that representation depends on consistency. When your name, your role, your category, and your description are the same across your own site, your structured data, your LinkedIn, the directories you appear in, and the way third parties refer to you, the representation is crisp and the model can use it without hesitation. When those signals conflict, when you are a "growth partner" in one place, a "marketing agency" in another, and something bespoke and unnamed on your own homepage, the representation is fuzzy, and a fuzzy entity is a risky thing to recommend.

Given a choice between a crisp entity and a fuzzy one, the model reaches for the crisp one. This is why a smaller, clearer competitor can be recommended ahead of a larger, vaguer you. Size is not a signal these systems read. Clarity is. A competitor that describes itself the same way everywhere, marks up its identity with structured data, and uses plain category language on its pages is handing the model a confident association. A company that is coy about what it does, that leads with a clever tagline instead of a clear statement of category, is handing the model uncertainty.

The fixes here are entity work, not content volume. Make your name, category, and description identical everywhere you appear. Use the correct Organization and Person structured data with consistent sameAs links to your real profiles. Say, in plain words, on the pages that matter, what you do and which category you compete in. None of this is about writing more. It is about removing the contradictions that make the model unsure it is looking at one company rather than three.

Reason three: their page answers the question and yours does not

The third gap is about extraction. Systems lift passages, not whole documents. When the model composes an answer, it is looking for a self-contained piece of text that addresses the question directly, and it will take that piece from whichever candidate page offers it most cleanly.

This is where a lot of well-written marketing pages fail. A page that sells beautifully, building a narrative across several hundred words before the point becomes clear, contains no liftable answer. The information is there, but it only emerges cumulatively, and there is no single passage the model can extract. A competitor whose page states plainly, near the top, what it does, who it is for, and how it is different, has handed the model exactly the sentence it needs. The competitor is not a better company. It is a more extractable one.

Consider the difference between a page that opens with a positioning line and a mission and a page that opens with "We help hospitality groups recover direct bookings lost to the major booking platforms, through a fixed twelve-week engagement." The second one can be quoted into an answer as it stands. The first one cannot. When the model is choosing which company to name and one page volunteers a clean, quotable claim while the other requires interpretation, the quotable page wins the mention.

The reassuring part is that the machine and the busy human want the same thing here. A buyer skimming your page also wants the answer up front, not after three paragraphs of throat-clearing. Writing for extraction is writing for the reader too. State what you do and for whom in a self-contained passage near the top. Use headings that describe the answer rather than tease it. Make each key claim stand on its own rather than depending on the paragraph before it. This is the rare case where optimising for the model and optimising for the person are the same edit.

Reason four: other sources vouch for them, not you

The fourth gap is corroboration, and it is the slowest to close and the hardest to fake, which is exactly why it carries weight.

Models treat what you say about yourself with appropriate caution. Every company claims to be a leader. What moves the needle is other, independent sources describing you as one. When your competitor appears in the "best providers of X" roundups, gets cited in articles about the category, and comes up in the forums and communities where your buyers gather, the model has corroborated evidence that this company belongs in the category, from sources that have no reason to inflate it. When the only place you are described as a category leader is your own homepage, the model has your word and nothing else.

You can see this in the search results these questions produce. Community threads and independent roundups appear repeatedly in the pages that assistants draw their candidates from. If your competitor is present in those places and you are not, the model is reading corroboration for them and silence for you. That gap is not something you can write your way out of on your own domain. It is earned off it.

Closing this gap is the patient, compounding work: earning genuine mentions, getting included in credible comparisons and roundups, being present and useful in the communities where your buyers actually spend time, and giving other people specific, quotable reasons to describe you as belonging to your category. It does not happen in a fortnight, and there is no shortcut worth taking. Attempting to manufacture it, through planted mentions or text that instructs a model to recommend you, does not work and risks your domain being treated as adversarial. The only durable version is the real one.

How to tell which reason is costing you

Four causes, and the wrong fix for the wrong cause is wasted effort. So do not guess. Diagnose.

Build a fixed set of prompts covering how buyers describe your category, the specific problems you solve, and your named competitors. Run each one in a clean session, with memory and personalisation turned off, across the assistants your buyers actually use, and run each one several times because these systems are non-deterministic and a single answer tells you very little. I have set out the full method in how to measure whether your brand appears in AI answers, and the same method is what turns this article from theory into a task list.

For each result, record four things, and each one points at one of the four reasons:

Are you retrieved or mentioned at all, even when the prompt is practically about you? If not, you are looking at reason one, and occasionally reason two.

When you are mentioned, how are you described, and is the description correct? If it is vague or wrong, that is reason two: the model is unsure who you are.

When a competitor is cited, is there a specific passage lifted from their page that you also cover, but only in prose the model could not extract? That is reason three.

Who else is named, and where do those companies show up across the wider web? If a competitor is consistently "the leader" and appears in the roundups and communities you are absent from, that is reason four.

The pattern across your prompt set tells you where to spend. Absent everywhere points at retrieval. Present but misdescribed points at entity clarity. Beaten on specific questions points at extraction. Beaten on general authority points at corroboration. Most companies have more than one gap, but they rarely have all four equally, and the diagnosis tells you which one is actually costing you the recommendation.

The takeaway

When an assistant recommends your competitor and not you, it is not scoring your companies against each other. It is naming the one it has the clearest, most corroborated, most extractable evidence for. That is a gap you can find and close, in order: get retrieved, be a clear entity, answer the question on the page, and earn the corroboration that you cannot write yourself.

The one step most teams skip is the first: they assume they know which reason is costing them, redesign a page or rewrite their positioning, and never confirm whether it moved anything. It is all measurable. If you want that measurement done properly, with a fixed prompt set, controlled conditions, and a clear read on which of the four gaps is yours, an AI search visibility audit is where that starts.

This article is part of the SEO in the AI Era: The Complete Guide guide.