Why does ChatGPT recommend your competitor instead of you?

ChatGPT recommends a competitor instead of you when their pages are in the retrieval set it built for that question and yours are not, or when yours are in it but nothing on them can be quoted as an answer. Those are the only two places the failure can sit, and separating them is the whole diagnosis.

Retrieval comes first. The engine turns one buyer question into several searches, pulls a set of candidate pages, and only then writes. If your page never enters that set, nothing about your product, pricing or quality is being judged at all. The question was answered without you in the room.

Selection comes second, and it is where most B2B companies actually lose. AirOps' 2026 retrieval study followed 548,534 pages ChatGPT retrieved across 15,000 prompts and found only around 15% became a visible citation, which means roughly 85% of what the engine reads never reaches the reader. A page can be retrieved, read and dropped because the competitor's page said the thing more plainly.

The third possibility is narrower and worth ruling out early: you are cited but not recommended. The answer links your page as background and names somebody else as the vendor to consider. That pattern usually means your page explains the category well and never states what you sell, to whom, and where you are the right choice.

What this is not is a ranking penalty. There is no position to lose and no manual action to appeal. The retrieval set is rebuilt for every prompt, which is the bad news for anybody hoping for a stable position and the good news for anybody currently absent from it.

How do you find out which sources the AI answer was built from?

You find the sources by reading the citation list under the answer and logging it, which is the only part of the process the engines expose directly. Ask the buying question the way a buyer would phrase it, then record every company named and every URL cited, with the date and the engine.

Run each question several times before drawing a conclusion. The source set rotates between runs for reasons that have nothing to do with your site, so a single answer is a sample rather than a finding, and our guide on why AI engines give a different answer every time covers how much movement is normal.

Then sort the cited URLs into three buckets: your competitor's own pages, third-party pages that name them, and pages that name nobody. The mix tells you what kind of problem you have. A list dominated by round-ups and review sites is an off-site problem. A list dominated by the competitor's own guides is a content problem on your domain, and it is the one you can act on this month.

Google rankings are a weak proxy for this and should not be used as one. Ahrefs' 2026 overlap analysis, run by data scientist Xibeijia Guan across 15,000 long-tail queries, found only about 12% of URLs cited by ChatGPT, Gemini and Copilot ranked in Google's top 10 for the same prompt. Checking your rankings tells you about a different system.

Write the log down in a form you can compare next month. A screenshot proves a moment and a spreadsheet proves a trend, and the question of which questions to put in it is covered in how many prompts you need to track AI visibility.

What is actually different about the competitor page that gets cited?

The competitor page that gets cited is usually written in the words the buyer typed, and that single difference outweighs everything else measurable on a page. Discovered Labs' analysis of 2 million AI citations crawled 10,000 cited pages and found prompt-content alignment carried a standardised effect of +0.37, roughly three times the next strongest signal they tested.

The same study puts the popular checklist items in perspective. An FAQ block measured at +0.07 in the same analysis, which is real and small next to the alignment figure. A competitor is rarely beating you because they added schema. They are beating you because their page answers the question in the question's own vocabulary.

Read the two pages side by side against one prompt and the gap is usually obvious within a minute. Their heading is the question a buyer typed and their first sentence answers it, while yours is a product name followed by three paragraphs of context. Their page also tends to name the use case, the company size and the constraint the buyer has in mind, where yours describes capabilities and leaves the reader to work out whether any of them apply.

The second difference is quotability. An engine lifting a sentence needs a sentence that stands alone, which means a direct claim with its own subject rather than a pronoun pointing at the paragraph above. Pages built from self-contained sections survive being read out of order, and pages built as a narrative do not.

What to compare	What the cited page usually does	How to check yours
Vocabulary	Uses the buyer's nouns and units, not internal product names	Search the prompt and read the words in the top results and the People Also Ask box
Heading shape	Headings are the questions buyers ask, worded as asked	Read your H2s aloud and ask whether anybody would type them
First sentence	Answers the heading outright before any context	Delete the first sentence of each section and see if the answer is still there
Sourced claims	Statistics carry a named source in the same paragraph	Count unsourced numbers on the page
Scope statements	Says who the product is for and who it is wrong for	Look for a sentence naming a buyer type and a constraint
Crawler access	Readable without JavaScript, not behind a form	Fetch the page with scripts disabled and read what remains

Is your competitor named because their domain is stronger than yours?

Domain strength is part of it and it is not the part you can do anything about, which is why the diagnosis should not stop there. Discovered Labs' analysis found a domain's perceived authority was around six times more influential than the strongest individual page-level feature, so a well known competitor does start ahead.

The useful finding sits underneath that one. In the same work, prompt-content alignment was the page-level signal that held up as the strongest lever available on a page, which means the question is not whether you can beat their domain but whether you can win specific questions on specific pages. Broad category prompts favour the famous. Narrow, qualified questions do not.

This is why the winnable ground is usually one level down from the head term. A prompt naming a buyer type, a constraint, an integration or a region has a smaller retrieval set, and the pages competing for it are fewer and often worse. Our guide on whether a small site can get cited by AI works through where that leverage is real and where it is not.

Treat an unknown competitor winning a prompt as information rather than an insult. If a firm with less brand recognition than yours is being named, domain authority is demonstrably not what decided that answer, and whatever did decide it is on the page. That is the cheapest research available in this whole exercise.

Where the gap genuinely is off-site, the work changes shape rather than disappearing. Round-ups and review sites appear in a lot of shortlist answers, and getting into them is outreach with a different skill set, which getting into the best-of lists AI engines quote sets out.

How do you tell a real gap from an answer that changes every time?

A real gap shows up as a pattern across repeated runs of a fixed question set, and ordinary answer variance shows up as a different company winning on each run. The test is repetition, and there is no shortcut that avoids it.

Run each question at least three times in a sitting and then again a week later, logging the named companies each time. A competitor that appears in nearly every run holds the question. A competitor that appears in one of six runs is inside the noise, and treating that as a crisis sends your content team after a problem that does not exist.

Keep the question wording fixed between runs. Changing the phrasing changes the retrieval set, so a question edited mid-measurement resets the series and makes every earlier run uncomparable. Write the question list down before the first run and leave it alone.

Separate the engines in the log rather than averaging them. The same prompt produces different source sets on different engines, and a competitor who owns the question on one may be absent from another, which changes where the work should go. Our guide to which AI engine to optimise for first covers how to pick when they disagree.

If the whole picture needs establishing before you can argue about it internally, a free visibility assessment runs your buyer questions across ChatGPT, Claude, Perplexity and Google AI Overviews and reports which companies are named and from which pages, which is the baseline every later comparison is read against.

What do you fix first when a competitor keeps winning the shortlist?

Fix retrievability first, alignment second, and everything else after that, because the order follows where a page can fail rather than what is easiest to ship. A page the crawler cannot read cannot be helped by better writing.

The technical checks are yes or no questions and they are cheap to answer: can the crawlers reach the page, does the content survive without JavaScript, is the heading order intact, is anything important behind a form. Our walkthrough on how to check if AI can read your website runs through each one.

Alignment is the second pass and the one with the most upside. Rewrite the page so its headings are the questions buyers ask and its first sentences answer them, using the nouns and units that appear in the live results for that prompt rather than the ones your product team prefers. This is a content job, not an engineering one, and it is usually done on pages that already exist rather than new ones.

Grounding is third. Every statistic on the page needs its source named in the same paragraph, because a number with no visible source reads as invented to a reader and gives an engine nothing to attribute. The research base supports the effort here: the KDD 2024 paper by Aggarwal and colleagues on generative engine optimisation, published as GEO: Generative Engine Optimization, measured visibility gains of up to roughly 40% from adding statistics, quotations and cited sources.

What should come last is the thing most teams start with, which is publishing more. New pages are the right answer for questions nothing on your site covers and the wrong answer for questions where you already have a page that nearly works. Fixing the page that was retrieved and dropped is faster than earning a new one into the set.

How long does it take to displace a competitor in AI answers?

Displacing a competitor takes weeks rather than days, and the first thing that moves is retrieval rather than recommendation. A rewritten page tends to start being read again before it starts being named, so the early signal is your URL appearing in citation lists at all.

Expect the sequence to be slower on broad prompts and faster on narrow ones. A qualified question with a small retrieval set can change hands after one good page, while a category head term with a dozen entrenched round-ups behind it may not move for a long time, whatever you publish. Our page on how long AEO takes to work sets out the realistic timeline.

Hold the measurement steady while you wait. Structural changes take weeks to register, so a run a fortnight after a fix mostly tells you whether the page is being retrieved again, and reading it as a verdict on the rewrite produces the wrong conclusion in both directions.

Plan for the position to be lost again, because citation sets rotate even when nothing on your site changes. Watching for a page that was cited and stopped is a standing job rather than a one-off check, and measuring AI citation decay covers how to catch it before the quarter ends.

The commercial stakes justify the patience. G2's 2026 research, a survey of 1,076 B2B decision makers reported in its announcement of the findings, found 69% of buyers chose a different software vendor than they had originally planned on the strength of AI chatbot guidance, and about a third bought from a vendor they had not previously heard of. The answer your competitor is winning changes shortlists rather than only impressions.

What else do people ask when AI recommends a competitor?