Ask an assistant to recommend a company in your category. Watch it name three, and watch yours not be one of them.
The reflex is to blame the website. Then someone points out that your competitor's site is worse than yours — slower, uglier, less content — and the theory collapses.
Here is what is actually happening, and it is a structural insight that changes where you spend money:
A search engine ranks pages. A model assembles an entity.
Google's job was to return a list of documents, and the unit of competition was a URL. A model composing an answer is doing something different: it is forming a compressed internal picture of a thing in the world — your company — from every source it has seen, and then describing that picture. Your website is one input. It is frequently not the loudest one.
The evidence for this is blunt. Ahrefs measured the overlap between what ChatGPT cites and what Google ranks and found that 83% of the URLs ChatGPT cites do not appear in Google's results at all, with only 6.82% matching Google's top ten — their AI search statistics collection. A companion study of the 1,000 pages ChatGPT cites most found 28.3% have zero organic Google visibility: no keywords, no rankings, quoted anyway — Ahrefs' ChatGPT citation study.
If your entire visibility strategy is "rank on Google", you have optimised for a surface that accounts for a small minority of what the largest assistant actually quotes.
So the question is not "how do I rank". It is: what does the web collectively believe about my company, and is it correct, consistent and recent?
The four failure modes
In practice, entity problems come in four shapes. Most companies have at least two.
1. Thin
There is not enough about you, anywhere, for a model to say anything confident. One website, a couple of social profiles, no independent mentions. A model asked to recommend will choose the entity it knows more about, because a confident answer is what it is optimising for.
This is the most common problem for good companies with quiet marketing. You are not being rejected. You are being skipped for lack of material.
2. Contradictory
Your website says one price, a marketplace listing says another, an old directory has an address you left in 2023, and a comparison site describes a service you stopped offering. Nothing here is malicious — it is simply unmaintained.
Contradiction is worse than silence. Given two conflicting figures, a model's safest behaviour is to describe you vaguely or to pick the competitor whose numbers agree with themselves everywhere.
3. Stale
The corpus's most detailed description of you is three years old. Freshness weighs heavily — 76.4% of ChatGPT's most-cited pages had been updated within the previous 30 days — and it compounds: a stale page is both less likely to be used and more likely to be wrong when it is.
4. Miscategorised
You describe yourself in language nobody searches in. You call it "digital transformation partnership"; the buyer asks about "a booking system for a clinic". If your own material never contains the buyer's words, you are not in the candidate set for the buyer's question — no matter how good you are.
5. Split
The one nobody looks for, and it is more common than any of the above.
You are not one entity in the corpus. You are two or three weak ones, because at some point the business existed under more than one name: a legal entity that differs from the trading name, a rebrand where the old site still resolves, a second domain bought for a campaign, a product name that outgrew the company name, a Cyrillic spelling and a Latin one that were never linked.
Each fragment collects its own mentions, its own reviews, its own half of the evidence. None of them accumulates enough to be the confident answer, and a model has no reliable way to know they are the same company unless the web tells it so explicitly.
Symptoms: an assistant describing your old positioning; a competitor named instead of you for your own product; your reviews appearing under a name your customers never use.
The fix is mechanical and worth doing before anything else on this list — consolidate to one canonical name and domain, 301 the rest rather than leaving them live, use the same name in every listing, and make sure the alternate spellings appear on the canonical site so the connection is stated somewhere a machine can read it. On Bronjoy this is why search matches across both Latin and Cyrillic: in that market the same business name is written both ways, and treating them as two things would have split every venue in the catalogue in half.
The listings half that nobody owns
Here is the part that surprises people, and it comes from running a directory ourselves.
Bronjoy listed 23,727 venues across Uzbekistan in September 2026. Most of those businesses had never built a website. Their entire machine-readable presence — name, category, address, coordinates, hours, services, price range — was a listing someone else wrote about them.
That is not unusual. For a large share of local and mid-sized businesses, third-party listings are the majority of what any model knows. And nobody inside the company owns them, because listings sit in the gap between marketing (who thinks it is IT), IT (who thinks it is marketing) and operations (who moved office and told the landlord, not the internet).
Third-party sources also punch above their weight in the answers themselves. Ahrefs found user-generated platforms cited more often than their share of the web would predict — Reddit appearing 3.4 percentage points more than expected, Quora 2.2, YouTube 2. Places where other people talk about you are not a supplement to your website. They are a parallel corpus with its own weighting.
When we built those venue pages we treated the data as a product: server-rendered pages per city and category, LocalBusiness structured data with addresses, coordinates and ratings inside ItemList collections, breadcrumb and FAQ markup, and a slug history so that a renamed venue's old URL still resolves to its current page. That last detail exists because a broken link is not just a lost visitor — it is a fact about you disappearing from the corpus.
The same discipline shows up in localisation. On Bronjoy, English text appearing on a Russian or Uzbek page is treated as a bug, and directory vocabulary — services, amenities, payment methods, landmarks — is translated once into a shared dictionary rather than stored per venue. A business described in three inconsistent vocabularies is three weak entities instead of one strong one.
Sources decide verdicts, not just facts
One more thing worth knowing about how these answers get made.
On Finkonkurs we build profile pages for the people behind crypto projects, assembled from AI research dossiers. One rule in that pipeline is deliberately counter-intuitive: government sanctions are recorded as neutral information, not as a scam signal.
Why? Because a research pipeline that treats every negative-sounding datum as evidence of wrongdoing will manufacture a verdict from context. The fix is to separate fact from judgement at the point of extraction.
The lesson for your business is the mirror image. If the loudest third-party material about you is a complaint thread, an outdated review or a competitor's comparison page, a model has no mechanism for weighting it fairly. It has facts, and it builds a verdict. You cannot appeal. You can only change the inputs.
A consensus audit you can run in an afternoon
Five checks. No tools required.
1. The name test. Ask four assistants "what is your company?" Record the answer verbatim. Mark every claim as correct, outdated or invented.
2. The category test. Ask the buying question, not the brand question — "who builds booking systems for clinics in your city" — and see whether you are in the candidate set at all. This separates thin from miscategorised.
3. The consistency sweep. List every place your name, address, phone, price range and service list appears: your site, Google Business, Apple Maps, industry directories, marketplaces, old press, partner pages. Compare them in a spreadsheet. Fix the contradictions before you write anything new.
4. The freshness pass. Find the five most detailed third-party descriptions of you. Check their dates. If your most authoritative external description predates your current offering, that is your highest-priority fix.
5. The vocabulary check. Take the twenty phrases your customers actually use. Search your own site for each. Count how many return nothing.
The output is a ranked list, and it is almost always cheaper to fix than a redesign.
Sequence it, because the order changes the cost
The audit produces a list. The order you work it determines whether it takes a fortnight or a year.
Week 1 — stop the contradictions. Reconcile name, address, phone, price range and service list everywhere they appear. This is pure data hygiene, needs no new content, and it is the only step that makes every later step more effective. Contradictions actively suppress you; removing them is the highest-leverage hour on the list.
Week 2 — fix the split, if you have one. Consolidate to one canonical name and domain. Doing this after you publish new content means publishing it to the wrong entity.
Weeks 3–6 — answer the buying questions. One question, one page, a number in the first two sentences, a date at the top. Start with the five questions that come up most in sales calls, because you already know the answers and they are already written down in your email.
Ongoing — maintain the dates. A quarterly pass that updates figures and re-stamps the date is worth more than a new article, given that 76.4% of ChatGPT's most-cited pages had been updated within 30 days.
Never — buy mentions. Paid placement in listicles and "top 10 agencies" pages is the obvious shortcut and it degrades fast, because those pages are transparently commercial and the assistants are getting better at discounting them. It also puts your entity's description in the hands of someone whose business model is selling the same slot to your competitor next quarter.
Where the work lands
Roughly half of this is content and copy — being specific, using the buyer's words, publishing facts instead of adjectives, keeping dates true. We cover the site side of that in AI search is not a traffic channel, and why the shortcut everyone reached for does not help in llms.txt will not get you cited.
The other half is operations: owning your listings, reconciling your own numbers, and making sure that when something changes in the business, the change propagates to every place a machine can read it. That is not a campaign. It is a maintenance habit, and the companies that keep it will look, to a model, like the ones that exist most clearly.
If your site is the weak link, we build high-conversion websites that are fast, server-rendered and structured from the first commit — 3–5 days for a landing page, from $99. If the site is fine and the problem is that it says nothing quotable, we will tell you that instead. Saying so is one of our five commitments.
Tell us what you need. You get a fixed price and a delivery date back — free, and with no obligation to proceed.
Frequently asked questions
Can I just get more backlinks?
Links help with classic search and they do carry some weight here, but the mechanism is different. What matters is whether independent sources describe you accurately and consistently — a detailed, current mention with no link often does more than a link with no context.
How long does it take to change what a model says about me?
Weeks to months, and it is uneven across assistants because they refresh on different cycles. Correcting a contradiction usually shows up faster than building presence from thin, because you are replacing a bad signal rather than creating a new one.
Should I pay a reputation management company?
Be careful. Most of the work here is unglamorous data hygiene you can do internally: listings, consistency, dates, vocabulary. If a proposal is mostly about publishing new content and not at all about reconciling what already exists, it is solving the wrong half.
Does this apply to B2B as well as local businesses?
Yes, with a different mix. B2B entities are thin far more often than they are contradictory, because the buying questions are niche and there is simply less material. Detailed, specific case studies are usually the highest-leverage fix.
What if a competitor's comparison page is what gets cited about us?
Then the answer about you is being written by someone with an interest in your losing. The counter is to publish the comparison yourself, honestly, including where you are not the right choice. A page that concedes something is far more likely to be treated as a reliable source than one that does not.
