Finding a competitor recommended by AI while your own brand goes unnamed is now one of the most common reasons a marketing team commissions a visibility review. The usual response is to search for a tool that promises to fix it. That instinct is understandable, but it skips the part that actually determines the outcome, which is working out why a model can describe one business in your category confidently and cannot describe yours. This article sets out the diagnostic sequence a practitioner follows, in order, from confirming the pattern to separating the problems that can be influenced from the ones that cannot.
Why one competitor keeps getting recommended and you are not
When someone types "best cyber security company for a UK law firm"into ChatGPT, Gemini or Perplexity, the answer is assembled from two ingredients. The first is retrieval: the system runs searches, pulls live pages, and grounds parts of its answer in those sources, which is why citations appear at all. The second is parametric memory, meaning what the model already holds about the category from its training data, including how businesses in that sector are usually described and which names recur alongside which claims. Retrieval explains why answers change week to week. Parametric memory explains why some brands appear even when they are not cited, and why a brand that has recently improved its website can still be missing from the shortlist.
Recommendation is a confidence decision, not a ranking
Marketers are used to positions, so the temptation is to treat a named competitor as the equivalent of a number one result. Recommendation works differently. A model names the business it can describe in one sentence without contradicting itself, because naming a company carries an implicit assertion that the company exists, does the stated work, serves the stated market and is a defensible suggestion. Where the available evidence about a business is thin, conflicting or ambiguous, the safer output for the model is to leave that business out and name one it can characterise cleanly. That distinction matters because it changes the repair work: the objective is to remove ambiguity, not to climb a list.
The causes that show up most often in practice
Four patterns account for a large share of cases. Miscategorisation is the most common and the most fixable: the business is present in the model's view of the world, but filed under the wrong service line, the wrong sector or the wrong region, so it never surfaces for the prompts buyers actually type. Thin third-party coverage is next, where almost everything a model can find about the company comes from the company itself. Inconsistent naming follows, including trading names, legal entity names, abbreviations and rebrands used interchangeably across the website, Companies House, directories, LinkedIn and trade press. Contradictory descriptions complete the set, where the homepage claims one specialism, an old landing page claims another, and a directory profile lists a service that was retired two years ago.
Why "we rank number one on Google"does not transfer
Strong organic rankings help, but they do not carry across cleanly. Independent analyses of Google's AI Overviews published through 2024 and 2025, including work by Authoritas and by Ahrefs, have reported that a meaningful share of the URLs cited in AI Overviews do not appear in the top ten organic results for the same query. A brand can hold page one for its main commercial terms and still be absent from the answer that shapes the buyer's shortlist, because the sources feeding that answer sit elsewhere: a trade publication, a review platform, a niche directory, a forum thread, a supplier listing. Classic SEO remains useful groundwork. It is not a substitute for checking what the models actually say.
Confirm the pattern before reacting to a single screenshot
A single screenshot of a rival being named proves very little. Answers vary by platform, by session memory, by phrasing, by location and by the order of the conversation, so one chat is an anecdote rather than a measurement. The reliable unit is share of recommendation across a weighted set of prompts, tracked over time.
Rerun the question across platforms and phrasings
The same question should be run across ChatGPT, Gemini, Perplexity, Claude, Copilot and Meta AI, because each ecosystem appears to weight sources differently and a brand that is invisible in one may be present in another. Phrasing then needs to move the way buyers move: "best", "most trusted", "cheapest", "alternatives to [competitor]", "who should I use for", plus location and sector qualifiers such as Manchester, Leeds, Bristol or "for UK SMEs". Price-led prompts are worth testing separately, since they tend to favour companies that publish figures openly. A brand can hold a solid position on "best"prompts and disappear entirely on "alternatives to"prompts, which is exactly where competitive displacement happens.
Personalisation makes quick checks misleading
Logged-in sessions with memory enabled will often reflect earlier conversations, saved preferences and inferred location, so a marketer checking their own brand in their own account is not seeing what a buyer sees. Clean sessions, incognito windows and, where available, memory turned off give a more neutral reading. Answers may still vary between two clean sessions run minutes apart, which is normal and is itself an argument for repetition over single tests.
Build a prompt set and monitor it
A workable prompt set for one brand typically runs from around 15 questions for a narrow local service to 100 or more for a multi-service business across several markets, weighted towards real purchase questions rather than vanity phrases such as "is [brand] any good". Daily monitoring across that set produces a share of recommendation figure that can be compared week to week, which is the basis of any sensible attempt to check how visible your brand currently is in AI answers before making changes. Spot-checking produces noise. Repetition produces evidence.
Trace the sources behind your competitor's mentions
Once the pattern is confirmed, the next step is source attribution: reading the citations, links and named references in the answers where the rival appears, and recording which publishers, directories, review platforms, trade titles and forums the model leaned on. Perplexity and Copilot make this straightforward because sources are listed. ChatGPT and Gemini are less consistent, so the practical method is to run the same prompt several times and record every source that appears, then look for the publishers that recur regardless of phrasing.
Source weighting is category-specific
The advice to "get on more listicles"is close to useless until the specific listicles are known. The publishers and directories that carry weight for hotels are not the ones that carry weight for law firms, and neither set resembles what surfaces for cyber security vendors, PR agencies or universities. Regulated and professional sectors tend to draw on register-style and credential sources, hospitality draws heavily on booking and review platforms, and technology categories often draw on comparison sites, analyst-adjacent blogs and community threads. Reddit and other user-generated sources have been prominent among cited sources in reporting through 2024 and 2025, but the share attributed to any single platform has moved as models, licensing arrangements and ranking systems change, so a source strategy built on one community is fragile by design.
The output of this stage is not a theory. It is a shortlist of named pages and publishers where the competitor appears and the brand does not, ordered by how often each source turns up in answers to the prompts that matter commercially.
Diagnose your own brand against the six things AI needs to name you
The same six-pillar checklist can be applied to a competitor's public footprint and to your own, which is what makes it useful for structured AI competitor benchmarking and ongoing implementation rather than guesswork.
Clarity. Can a model state in one sentence what the business sells, to whom, and where, using only publicly available text. If three people in the company would answer differently, a model will not answer confidently.
Consistency. Same legal and trading name, same categories, same locations, same core claims across the website, Google Business Profile, LinkedIn, directories, press coverage and any partner listings.
Trust. Independent corroboration rather than self-description: credentials and certifications, named people with real biographies, verifiable case detail, and reviews with enough volume and recency to be treated as signal.
Visibility. Whether the pages that answer buyer questions exist at all and are discoverable through ordinary search, since retrieval usually starts there.
Freshness. Whether service lists, pricing, team pages and location details reflect the business as it operates now, with dates that support that.
Technical foundations. Whether pages render server-side, carry sensible structured data, and are reachable by AI crawlers rather than blocked in robots.txt or behind aggressive bot protection.
Scoring each pillar produces a composite AI visibility position that can be retested after changes, against a baseline captured before anything was touched. Without that baseline, improvement is a matter of opinion.
Fixes that change AI answers, ranked by how fast they move
Quick corrections
Wrong category, an outdated service list, a mismatched trading address, a stale company description on a high-visibility profile: these are cheap to fix and frequently produce the fastest change in answers. DarkInvader, a cyber security brand, was found to be categorised incorrectly in AI answers, which was addressed through implementation work and then verified by retesting against the original baseline rather than assumed to have worked. Miscategorisation is the pattern worth looking for first in any diagnosis, because correcting it can move answers faster than a content programme that runs for months.
Medium effort
Comparison and alternatives pages written honestly, pricing that states real figures or real ranges, and FAQ content phrased to match the exact questions buyers put to assistants. Models handle specific, self-contained answers better than brochure prose. A page that says what the service costs, who it suits and who it does not suit gives a model something safe to quote.
Slow but durable
Earning coverage on the third-party sources the models actually pull for your category, and building review depth on the platforms those answers reference. This work takes months and cannot be shortcut, which is precisely why it is defensible once established.
What not to do
Mass AI-generated pages dilute clarity and add contradictions, which is the opposite of the goal. Fake or incentivised reviews are a compliance risk in the UK under consumer protection rules and are increasingly detectable. Blocking AI crawlers to protect content is a self-inflicted visibility problem: if a model cannot retrieve the site, it will describe the business from whatever third parties say instead, which is usually the situation that created the competitor recommendation in the first place.
AI competitor benchmarking: what to track and how often
Share of recommendation across the weighted prompt set is the headline metric, expressed as the percentage of prompts in which the brand is named, alongside the same figure for each tracked rival. A single position on a single prompt is a vanity number. Weekly category leaderboards, of the kind produced for sectors such as hotels, restaurants, law firms, cyber security, PR agencies, universities and event security, are useful for spotting movement: a rival gaining ground usually shows up first on a narrow cluster of prompts, often "alternatives to"or price-led phrasings, before it spreads.
Before-and-after retesting is the discipline that separates evidence from anecdote. Capture the baseline, change one set of things, retest the same prompts on the same platforms, and record what moved and what did not. Some changes will produce no visible effect, and saying so plainly is part of the work.
Larger organisations add complications: multiple brands under one group, several markets with different competitor sets, and multilingual prompts where the same question in German or French returns a different shortlist. Governance matters at that scale, and the four functions of the NIST AI Risk Management Framework (Govern, Map, Measure, Manage) provide a reasonable structure for documenting who owns AI visibility reporting, what is measured and how claims are checked. It also imposes an honest limit: no provider controls what happens inside a third-party model, and any supplier promising guaranteed placement inside ChatGPT is overstating what is possible. Organisations with several brands or markets can discuss enterprise monitoring requirements directly.
When to fix it yourself and when to bring in help
A one-week self-check is genuinely feasible with free accounts and a spreadsheet. Write 20 buyer prompts, run them in clean sessions across ChatGPT, Gemini, Perplexity and Copilot on three separate days, and record for each run which brands were named and which sources were cited. Then audit your own footprint against the six pillars and list every contradiction found. A week of this work commonly surfaces fixable errors, typically a stale profile or a category mismatch.
Outside help is worth considering when answers contradict each other across platforms, when a specific wrong fact persists after corrections have been made, when the sector is regulated and inaccurate claims carry legal exposure, or when the prompt set is too large to run manually. Ongoing work of this type usually combines an audit, competitor benchmarking, source analysis, content and technical recommendations, implementation and repeat measurement, with results evidenced by retesting against the original baseline rather than asserted. Published options and inclusions are listed on the AI visibility plans and pricing page, and pricing should be checked there at the time of reading.
The practical point to hold on to is this: a competitor recommended by AI is a symptom, and the diagnosis is almost always clarity, consistency, corroboration or retrievability rather than luck. Confirm the pattern across platforms, trace the sources feeding the answers, test the six pillars against your own footprint, fix the errors first, then measure again.
Frequently Asked Questions
Why does ChatGPT recommend my competitor instead of my business?
Usually because the model can describe the competitor confidently and cannot describe your business without risk of contradicting itself. The common causes are miscategorisation, thin independent coverage, inconsistent names and claims across the web, and pages that cannot be retrieved. Recommendation reflects the model's confidence, not a ranking you can climb directly.
How can I check whether a competitor is recommended by AI more often than us?
Build a set of 15 to 100 buyer prompts, run them on the same days across ChatGPT, Gemini, Perplexity, Claude, Copilot and Meta AI in clean sessions, and record which brands are named each time. The comparison you want is share of recommendation across the whole set, not a single result. Repeat it on a schedule so movement can be distinguished from normal variation.
Can I get an AI assistant to correct wrong information about my company?
There is no direct edit function, so the practical route is to fix the underlying evidence: correct the category and details on your own site and on the profiles and directories the answers cite, publish clear current facts, and remove contradictions. Answers often update as those sources are recrawled, though timing varies by platform. Retesting the original prompts is the only way to confirm the change took effect.
How long does it take to start appearing in AI recommendations?
Simple corrections such as a wrong category or an outdated service list can show up within weeks, sometimes sooner. Work that depends on third-party citations and review depth is measured in months. No timeline can be guaranteed, because retrieval frequency and model update cycles sit outside any supplier's control.
Do AI answers differ between ChatGPT, Gemini, Perplexity and Claude for the same question?
Yes, and often substantially, because each system retrieves from different indexes and appears to weight sources differently. A brand may be named consistently by Perplexity and absent from Gemini for the same prompt. That is why any assessment worth acting on covers several platforms rather than the one the team happens to use.