Between what an AI engine answers and what a tool records as an answer sit texts that look the part: the prompt sent straight back, a statement of intent, a generic message unrelated to the question, a map credit. None of them makes a run fail, and four of the five cases below can create a mention that does not exist.
So before you defend a three-point gain in a board meeting, or move budget because of a drop, you need to be able to open the answers behind the number. The Nanga team observed these cases in its own data between July and October 2026. Six questions for any tool follow them.
One caution: these rates come from a single team, with no outside audit. As of 26 October 2026, no independent study measuring these errors across tools could be found. Nobody can say what a normal rate would be, or who handles this best.
What is a false visibility signal?
A false visibility signal is a text counted as an AI engine's answer when it is not one, or a mention counted when the engine did not name the brand in its answer.
It differs from two better-known problems. A factual error is a real answer that is wrong. Randomness is a real answer that changes from one ask to the next. Taking more measurements reduces randomness but leaves a constant share of invalid rows in place: 1.7% false is still 1.7% at any volume.
The simplest checks, non-empty text in the right language, do not stop it. Nearly all five cases below pass them.
What happens when an engine returns your own prompt?
The row can count as a mention, if the prompt names the brand. The Nanga team saw records whose answer was the prompt, word for word, with no source: about 0.6% of one engine's rows from late September 2026. It took two weeks to notice, because no error ever surfaced.
Take "Why trust Alba & Co?". Returned as is, that text names Alba & Co, no competitor, and passes for an answer. Only prompts that name the brand are exposed, which compounds a bias already present in brand-heavy prompt sets.
The filter has to stay strict. Real answers often open with the question as a heading, so only an exact match with the whole prompt marks an echo.
Why can a plan line count as a mention?
Because it names the brand before any search has run. "I will check public customer reviews of Alba & Co" contains no information, but the name is there. OpenAI documents this kind of message: its GPT-5 prompting guide, written for developers and now archived, says the model is trained to give an upfront plan and then progress updates.
The trouble starts when that sentence is all that is left. On ChatGPT, the Nanga team saw these rows go from under 0.1% to about 1.7% of answers from late September 2026, scattered across prompts, and nine days passed before they were detected.
"No source, so suspicious" does not work as a test. Schulte and co-authors (University of St. Gallen, April 2026, four verticals in Switzerland, German-language prompts, January to March 2026) report that 57.8% of ChatGPT runs had no citation at all. Discarding everything without a source would throw away many legitimate answers. A plan line is identified by several signs together: short text, no search and no source, no question to the user, a first-person plan.
What is a generic message stored as an answer worth?
Nothing, and it is the one case of the five that does not manufacture a mention: it manufactures an answer. One day in July 2026, an engine returned the same message, unrelated to the questions asked, in the project's language, across a whole daily batch. It was stored as valid because it was not empty and the language matched.
The effect depends on the tool's denominator. If the score divides by all answers, your visibility collapses on that engine although nothing changed for your brand. If it divides by answers that name a brand, the message drops out of the calculation and the score barely moves, but one engine has stopped being measured and the number does not show it.
How does interface text turn into a brand?
It happens when a tool reads what surrounds an answer as part of the answer. Maps are a good example: OpenStreetMap asks for its attribution notice to be displayed, and Mapbox requires its logo and a text attribution. In August 2026 the Nanga team saw map providers' credits extracted as brands. Likewise, a block of working steps placed before the answer was read as its start, and its lines name brands and sites the engine had merely looked at.
Cleaning out everything would be a mistake, though. Map text touched about 4% of one engine's answers across the weeks of September and October 2026. Once it was removed, detection of the tracked brand changed on none of the answers from active projects. The pinned places stayed, because an address placed on the map is a real recommendation from the engine. Frequent pollution does not prove a bias of the same size.
Is a brand name inside a URL a mention?
No, it is a source. When an engine writes "according to alba-co.com", it says where it read something, without naming you in its recommendation. An unbounded text search still finds "alba" in that domain and adds a mention. The Nanga team fixed this case in July 2026 after it inflated share of voice.
The line is blurrier than it looks. A strict rule can undercount a brand whose public name is its domain, while a short alias can collide with a country extension. Mentions and citations should stay in separate columns, because being cited without being named and being named without being cited call for different actions.
How many points can a single row move?
In the illustration below, between 2.5 and 4 points on a small topic; the same row weighs far less across a large program. Take a fictional topic of 10 prompts on 4 engines: 40 answers per run, of which 25 name at least one brand and 8 name Alba & Co.
| Event | Answers naming Alba & Co | Score on all answers | Score on answers naming a brand |
|---|---|---|---|
| Normal run | 8 | 20% | 32% |
| An echo naming Alba & Co replaces an answer that named only competitors | 9 | 22.5% | 36% |
| A generic message replaces one engine's 10 answers, of which 6 named a brand and 2 named Alba & Co | 6 | 15% | 31.6% |
Arithmetic illustration, with fictional figures and brand.
The same incident costs one score 5 points and the other 0.4: two honest tools can tell two different stories. Separately, at 1.7% of answers, a set of 50 prompts yields an average of 0.85 plan lines per run on the engine concerned (illustrative calculation), which only matters if the line names a brand.
The real risk lies in the shape of the curve. These artifacts begin on a specific date and draw a step, just like the effect of a campaign or a model change. A control group helps separate them from a real effect: a step that also hits untreated topics did not come from your work. What nobody can say is how far they moved the scores any given tool has published. The rates above describe rows, not score points.
Does an absence equal a zero?
Only if the engine really answered. Three situations need separate treatment:
- No overview served. Google says its AI Overviews appear only when its systems judge them additive to classic Search. The Nanga team found no overview on 11.7% of results over 14 days, on the prompts it tracks, which is not a market rate. Quietly dropping those rows would show 100% exposure everywhere.
- No brand named. An honest answer may name none. An internal check that equated "no brand" with "failure" flagged 42% of a sample of 500 rows, when under 1% had actually failed.
- Partial run. An incomplete analysis fed into a trend line looks like a market move. It should be flagged and kept out of the trend.
What can you check yourself in an hour?
Most of it, with no technical skills, as long as you can reach the answers:
- Sort one run's answers by length. Read the 50 shortest. Echoes, plan lines and generic messages show up quickly.
- Isolate sourceless answers that name only you. Some are legitimate, but this is where false positives pile up.
- Read the list of detected competitors. A map provider, an engine's name or a redirect domain gives away faulty extraction.
- Find answers where your name appears only inside an address. Count them separately.
- Date every break. If a score moves on the same day across all of one engine's topics, with no action on your side, suspect the measurement first.
What should you ask a GEO tool about its data?
A precise, dated answer beats a promise of reliability. These six questions apply to any vendor:
- Can I read the raw answer behind every number? Without it, nothing below can be verified.
- Which non-answers do you reject, with which recent examples? A team that has met the problem can describe it.
- How many real answers do your filters wrongly discard? A filter nobody has measured may just swap one bias for another.
- How do you separate mentions from sources? Ask about a brand whose name is a domain, and about a source wrapped in a redirect, which should resolve to the real publisher.
- What is your denominator, and what happens to an absence? The table above shows what is at stake. Look for "not measured" wherever 0% would be false.
- What happens to history and duplicates after a fix? A recalculated series, a flagged break or data left as is: each is defensible if announced. A second collection on the same day should not double your volumes.
What does Nanga do about these five cases?
Nanga's GEO module rejects prompt echoes, plan lines and recognized generic messages instead of counting them as answers. It removes identified interface text before analysis and treats a name inside a cited domain as a source. For projects tracking Google AI Overviews, questions with no overview leave the visibility metrics and feed an exposure rate. A partial or failed run is flagged, some screens show "not measured" rather than 0%, the AI Visibility calculation opens factor by factor, and the latest run's answers export as JSON, CSV or Markdown.
To read the answers behind your score, engine by engine: request a demo.
Frequently asked questions
What is a false AI visibility signal?
A text counted as an AI engine's answer when it is not one, or a mention counted when the engine did not name the brand in its answer. The Nanga team has observed five forms: prompt echo, plan line, generic message, interface text and a name inside a URL.
Do these errors always push the score up?
No. An echo or a plan line can add a mention, but a generic message creates none. It can pull down a score calculated on all answers while leaving a score calculated only on answers that name a brand almost untouched.
Is an answer with no source suspicious?
Not on its own. Schulte and co-authors (University of St. Gallen, April 2026) found that 57.8% of ChatGPT runs in their sample had no citation at all. A false signal shows several signs together: very short text, no search, no source, no information.
How can a non-technical team check a GEO tool?
Export the answers from one run, sort them by length and read the shortest. Then read the list of detected competitors: a map provider or an engine's name points to a problem. If the tool gives no access to its answers, its score cannot be verified.
Is there a benchmark rate for these false signals?
No. As of 26 October 2026, no independent study comparing these errors across tools could be found. The rates cited here, about 0.6% and 1.7% of one engine's answers, are the Nanga team's observations on its own data, with no outside audit.
Sources:
- Nanga team, observations on its own data, July to October 2026, with no outside audit.
- Julius Schulte, Malte Bleeker and Philipp Kaufmann (University of St. Gallen), Don't Measure Once: Measuring Visibility in AI Search (GEO), preprint, April 2026.
- OpenAI, GPT-5 prompting guide, archived and undated page, accessed 26 October 2026.
- Google Search Central, AI features and your website, accessed 26 October 2026.
- OpenStreetMap, Copyright and License, accessed 26 October 2026.
- Mapbox, Attribution, accessed 26 October 2026.