Hi everyone, this is Neo.

Let me start with an uncomfortable question: when was the last time you asked ChatGPT “what are some trustworthy brands in your industry?” or “which one is best for X?”

If you run an independent site, do yourself a favor today: ask an AI a question about your industry — without mentioning your brand name. Then read carefully how it “describes” you in the answer. If it even describes you at all.

Why? Because Duane Forrester — former Bing executive and current Search Engine Journal columnist — published a spine-chilling piece on September 3 titled When AI Has Nothing On Your Company, It Describes Someone Else. The thesis: when an AI has no material on your company, it won’t tell you it doesn’t know. It describes you as someone else.

The Real Danger Isn’t “AI Can’t See You.” It’s “AI Mistakes You for Someone Else.”

Until now, every conversation about AI search and independent sites has revolved around visibility: Does AI mention your brand? How many citations? What rank?

Forrester flips the table: the truly dangerous errors are the ones that look completely normal.

When a model has nothing on your company, here’s what it does:

It does not decline, it does not hedge, and it does not mark the spot where its evidence ran out. It reaches for the nearest well-documented thing — usually a competitor, a category average, or a version of you from three years ago — and it delivers that substitution in the same confident register it uses when it actually knows.

Sit with that for a second. The AI isn’t fabricating wildly; it’s filling a blank — with whatever statistically resembles you most. And from its own perspective, this isn’t even an error: it simply produced the most probable description of “a company like yours.”

But the user reading the answer doesn’t know that. Neither do you. Until the day you stumble on it: “wait, AI says we use our competitor’s pricing model?”

Why AI Would Rather “Improvise” Than Say “I Don’t Know” — This Isn’t Mysticism, It’s Mechanics

Forrester doesn’t stop at the phenomenon — he backs it with research, and I’ve distilled it into a clean causal chain:

Link one: models are structurally bad at “unpopular knowledge.” He cites the ACL 2023 paper by Alex Mallen and colleagues, which probed ten large language models on knowledge recall. The finding is brutal: models struggle specifically with less popular factual knowledge, and scaling mostly improves recall of popular facts while barely touching the long tail. Hoping the next model generation will just “know you automatically”? Not a chance. If training data about you is thin, a bigger model doesn’t help.

Link two: models recite from memory rather than reading what’s in front of them. He cites Longpre and colleagues’ EMNLP 2021 work on knowledge conflicts: models lean heavily on memorized (parametric) information instead of actually reading provided context. The practitioner question becomes: is it reading, or is it guessing? More often than not: guessing.

Link three: stack those two together and you get substitution. A question about you arrives; your entity is sparse in the weights while your neighbor (competitor, category template) is dense. The model drifts toward that dense neighbor — reliably, in the same direction every time — because that gradient is a property of the training distribution, not a random sampling glitch.

Forrester is emphatic that this is not what we usually call hallucination. Random hallucination is unstructured invention — a fake citation, a fabricated stat — and better models measurably reduce it. What he’s describing is systematic, directional, and predictable. Models are genuinely getting better at not inventing things. They are not getting better at knowing about companies nobody wrote about.

That knife lands in the heart of every small and mid-sized brand. Big brands have wall-to-wall media coverage; the model would have to work hard to get them wrong. Independent site sellers are, almost by definition, “companies nobody wrote about.”

The Four Shapes of Substitution — Each One Gets Misdiagnosed as Something Else

Forrester breaks substitution into four forms. Here’s each one mapped to an independent-seller scenario:

Form one: silent analogy — the closest well-documented neighbor stands in for you.

The model reaches for the nearest well-documented similar company and describes it as you. Classic failure: your pricing model becomes your biggest competitor’s pricing model; your implementation timeline becomes the “industry average” timeline. Nothing in the answer signals a swap happened — because from the model’s side, none did. It produced the most probable description of a company like yours.

Form two: staleness presented as currency — the you from three years ago.

The model holds a version of you that was accurate at some point and states it in the present tense: discontinued products still on sale, departed executives still on the team page, markets you exited still listed as served. Parametric memory carries no timestamp and no expiry warning.

Form three: thin evidence wearing thick confidence.

You have one trade-press write-up; your competitor has forty independent sources. In the AI’s answer, both sound identical — the model never surfaces how much material it’s standing on. A claim resting on a single blog post arrives sounding like consensus.

Form four: category knowledge applied to your name.

The model knows your industry well and you hardly at all, so it answers the industry question and stamps your name on it. This is the hardest to spot, because the answer is usually true about the category — it survives a casual accuracy check.

The Kicker: All Four Errors Slip Straight Through Your Content Audit

The section that hit me hardest — read this twice if you run an independent site:

You cannot find this by auditing your own content. Every content audit ever built inspects what exists: pages, structure, coverage, accuracy, freshness. This failure lives in what does not exist, in a place you do not own.

Get it? Your content audit comes back clean. Pages are fine, schema validates, coverage looks reasonable against the competitive set. And the model is still telling people things about your company that no page of yours supports and no page of yours could have prevented.

Why? Because the evidence that would have prevented each of these four failures was mostly written by other people, over years — not by you. No amount of inspecting your own material reveals that shortfall, because the shortfall isn’t in your material.

Can “just publish more content” fix it? Forrester pours cold water on that too. He cites Sciavolino and colleagues’ EMNLP 2021 work on entity-centric questions: dense retrieval underperforms sparse methods badly on entity-rich questions and generalizes reliably only to common entities. In plain language: the popularity gradient that thins out your presence in the model’s weights reappears in what retrieval brings back — retrieval is a second application of the bias, not a correction of it. The escape hatch most tactical advice is selling runs through the same narrow door as everything else.

The Data: How Often AI Gets You Wrong Is Worse Than You Think

Mechanism talk can sound alarmist, so here’s hard data from independent industry studies in 2025–2026:

Searchable’s 13,000-query study (July 2026): 13,000+ queries across ChatGPT, Perplexity, and Gemini about 165 London businesses, verified against official Companies House records. Result: 93% of companies had at least one key fact misstated or missing. Half of all SMEs received at least one outright fabricated fact — versus 32% of large companies. Small brands’ fabrication rate (5%) was 2.5x large brands’ (2%). Brand-name confusion hit 4% of SMEs versus 0.7% of large companies.

Insites’ AI visibility benchmark (10,000 US local businesses): ChatGPT “knew” 93.59% of businesses — reassuring, right? But cross-checked against Google Business Profile, only 55.58% of answers had every detail match. Most common errors: wrong phone number (30.09%), wrong website (15.59%), wrong business name (6.38%). And in 1.68% of searches, none of the details matched — the AI was confidently describing a rival or an unrelated company as you.

SOCi’s Local Visibility Index (340,000+ locations): Gemini scores 100% profile accuracy because it reads Google Maps data directly; ChatGPT and Perplexity, assembling third-party sources, land at 68% — roughly one in three answers carries an error.

And one number to sit with tonight: BrightLocal’s 2026 survey found 45% of consumers now ask AI for local business recommendations — up from 6% a year earlier. Users are handing “who do I choose” to AI, and the AI’s file on you may contain a single article nobody read.

A Hall-of-Mirrors Case That Should Chill You

There’s a case study in Forrester’s piece I strongly recommend reading in full — it elevates the whole problem.

In June, a vendor article aimed at brand teams circulated, warning that “AI hallucinates claims about your products.” It looked professionally researched: a quotation attributed to Percy Liang, director of Stanford’s Center for Research on Foundation Models; the Stanford HAI AI Index; Nielsen; Gartner; the IAB; MIT Sloan…

Forrester checked every citation. The Percy Liang quote doesn’t exist anywhere — not in any paper, talk, or interview. The paper said to classify brand hallucinations into four named types doesn’t appear in the open, indexed ACL Anthology. Every report title sounded exactly like a real one; every link pointed to an organization’s homepage and never to an actual document.

See the hall of mirrors? This is an article warning that AI invents confident claims about brands — and it invented confident claims about researchers and institutions to make its case. The author needed authoritative support for claims about brand-level hallucination rates; that support doesn’t exist in the volume the argument required. So the gap got filled with things shaped like evidence: real institution names, plausible titles, a correctly identified Stanford director, working links.

Then the marketing director read it, believed the numbers, and put them in a board deck.

The substitution moved from a model, to an article, to a person, to a decision — with no one along the chain doing anything obviously wrong. That’s what makes this terrifying: it doesn’t stay inside machines. It propagates.

So How Do You Even Detect It? Only by Flipping the Direction

Forrester’s diagnostic advice is the most actionable part of the whole piece:

You find substitution by watching outputs, not by inspecting inputs. The diagnostic surface is something you do not own and cannot fully sample.

Concretely:

First, change what you’re looking for. Don’t ask “is it accurate?” Ask: does every claim the model makes about you trace back to a source you or others actually published? A capability you don’t have, described in confident detail. A timeline you never quoted. A competitor comparison built on dimensions you never set. Each one marks a spot where the model needed evidence about you, didn’t have it, and produced something anyway.

Second, how you sample matters more than how much you sample. Asking a model to “introduce brand X” is the least informative test possible — your name in the prompt is itself a retrieval cue that drags whatever thin material exists straight into context. Substitutions surface on questions where your name is absent and the model has to supply it:

• Category questions with a qualifier: “which brands serve the North American market for X?”
• Comparison questions naming only your competitor: “which fits scenario Y better, A or B?” (you’re not in the question — see if you appear as a third option)
• Questions about a capability rather than a company: “are there tools/brands that support feature Z?”

Third, run that repeatedly across the questions your buyers actually ask — because substitutions move. Treat the results as a map of where third-party description is missing, not a to-do list of pages to fix. Those are different documents.

Full disclosure: Forrester notes at the end that he runs an AI visibility measurement company, so he’s vested — judge accordingly. But the core argument stands on its own: a discipline built entirely around auditing what you own is structurally blind to a failure that happens entirely outside what you own. The cleanest content audit you’ve ever run tells you nothing about this. It never did.

Neo’s Take: What Independent Site Sellers Should Actually Do

What this article taught me is that the GEO homework we’ve been assigning is directionally right but missing the sneakiest step. We’ve been saying “make sure AI can see your brand,” then “make sure citations come from third parties.” Forrester reminds us there’s a checkpoint after visibility: once AI does see you, is the “you” it describes actually you?

For independent sellers, here’s how to turn this into four concrete actions:

First, run an “AI mirror test” today (half a day). Using the method above, ask 20–30 real buyer questions without your brand name across ChatGPT, Perplexity, Gemini (add Claude if you can). Log every description of you, flagging three types: descriptions that are clearly your competitor’s, information frozen a year in the past, and confident claims you never made. That log is your substitution baseline.

Second, build an inventory of third-party sources. Cross-reference the mirror test against the “neighbor” the AI borrowed. Then audit what external material about you actually exists: Google Business Profile, Amazon and marketplace listings, industry directories, review sites, forums, press coverage. Fill whichever cell is empty. Remember Insites’ conclusion: AI visibility is built on corroboration, not on any single perfect asset. The more independent places your data appears — consistently — the harder it is for a model to describe you as someone else.

Third, stop treating content volume as the cure. Publishing more pages helps, but it doesn’t cure the “long-tail sparsity” disease — retrieval amplifies the bias a second time. External mentions and consistency are the real medicine. This is exactly why I keep telling independent sellers to move budget from “writing more articles in a silo” to “getting mentioned out there”: influencers, industry media, Q&A communities, podcasts. Let third parties say who you are — don’t just say it yourself.

Fourth, put “how does AI describe me” on your regular cadence. Like rankings and Search Console, run the same fixed question set against the major AIs once a month. Substitutions move, competitors publish new content, and the you-from-three-years-ago gets older every day. This isn’t a one-time project; it’s ongoing monitoring.

One last thought: AI doesn’t go quiet when it doesn’t know you. It just quietly swaps you for someone else. If you don’t care how AI describes you, it will care for you — in the way you’d least want. In this era, your brand’s “portrait” inside AI is becoming a storefront more important than your homepage. Don’t let that storefront hang your neighbor’s sign.

I’m Neo, and I write about independent site SEO. If you’ve ever caught an AI describing your brand as someone else, come back and tell me what you asked and what it said — real examples are worth a hundred theories.