# Ask an AI what it thinks of your company

> We asked three AI engines what they thought of our own agency. Two say they do not know us, the third invents four different companies. The five-minute test, and how to stop leaving a gap to fill.

- Original page: https://dotsland.com/en/ask-an-ai-what-it-thinks-of-your-company/
- Language: en
- Published: 2026-11-09T09:00:00+01:00
- Updated: 2026-10-14T00:27:03+02:00
- Section: Digital Marketing
- Publisher: Dotsland, https://dotsland.com/en/

---

## In brief

- On 14 October 2026 we asked three AI engines what they thought of our own agency, Dotsland, without giving them web access: four questions, twice each, 24 answers.
- No answer describes Dotsland. GPT-4o and Claude say they do not know us, 16 times out of 16. Gemini makes things up: a digital resources platform, a mobile game, nursery management software, a new-build property website.
- So the risk is not being absent, it is being mistaken for something else, and it depends on which engine your prospect uses.
- The test takes five minutes, with no tools, and tells you whether it is urgent.
- The fix is not to get the engine corrected, it is to stop leaving it a gap to fill.

Ask an AI what it thinks of your company, and the answer will depend on the engine. On 14 October 2026, asked about our own agency without web access, GPT-4o and Claude said they did not know it in 16 answers out of 16, while Gemini described four different companies that do not exist under our name. None of the three described Dotsland. A poorly documented brand is not ignored: it is either absent or confused with something else.

> **How we measured**  
> On 14 October 2026 we asked the four questions in this article, twice each and in French, through the programming interfaces of OpenAI (GPT-4o), Anthropic (Claude Sonnet 5.5) and Google (Gemini 2.5 Flash), without web search. Each answer was read and classified by hand: abstention, invention, attribution to another product, or accurate description. The 24 raw answers are kept with their model and time.

*The test and the answers from our first reading, in August 2026, in thirty seconds.*

## What do AIs say about a company they do not know?

Here is what three engines answered, on 14 October 2026, to four questions about our agency.

| Engine | “I don’t know” | Invented company | Name attributed to another product | Accurate description |
| --- | --- | --- | --- | --- |
| **GPT-4o** (OpenAI) | 8 | 0 | 0 | 0 |
| **Claude Sonnet 5.5** (Anthropic) | 8 | 0 | 0 | 0 |
| **Gemini 2.5 Flash** (Google) | 2 | 4 | 2 | 0 |
| **Total** | **18** | **4** | **2** | **0** |

Gemini’s four inventions have nothing in common (quotes translated from French): “a French platform dedicated to buying and selling creative digital resources”, on which it gives an “overall very positive” opinion; a mobile game attributed to a real publisher; “a management platform for early childhood care facilities”; and “a platform specialising in new-build property in France”, complete with a list of competitors. Each call produces a different company, always described with confidence.

The contrast with the summer is sharp. On 9 August 2026, on the same brand and also without web access, our readings counted six invented identities, including an unofficial IPTV provider and a web host “generally not recommended”, and neither Gemini nor Claude said they did not know us. In October, Claude abstains every time. Engines change from one version to the next, and in August we did not record the exact version of each model: this is a trend, not a like-for-like comparison. A reading is only valid for its date and its model.

## Why does an engine make things up rather than stay silent?

Three mechanisms combine, and none of them is a bug.

**1. A model completes, it does not look things up.** Without web access, it does not search for anything: it produces the most probable continuation of your question. Faced with an unknown name, the most probable continuation is not always an admission of ignorance; it is often a description that looks like the thousands it has seen.

**2. The name carries misleading clues.** A name made of a common word, a widespread suffix or close to an existing brand steers the fabrication. “Dots” suggests a connect-the-dots game, “land” a world or a theme park: Gemini followed both trails.

**3. Abstaining is learned behaviour.** Saying “I don’t know” is trained, and publishers do not train it to the same degree. Our measurement shows it: on the same questions, on the same day, two engines stay silent and a third makes things up.

## And when the engine has web access?

Our test deliberately cuts off web access, to see what the model “knows” about you. When an assistant searches before answering, what counts becomes what it finds online about you, and each engine has its preferred sources. Both situations coexist depending on the tool and the user’s settings: that is why the work covers both what your site declares and what others write about you, as we show in our [citability audit](https://dotsland.com/en/ai-citability-audit/).

## How do you run the test for your company?

Ask the following four questions on three different engines, twice each, and paste the answers into a dated document, with the engine’s name. That document is your starting point.

1. What do you think of *your brand*?
2. Is *your brand* reliable and recommended?
3. What are the best alternatives to *your brand*?
4. What is *your brand* known for?

Then sort what you read into one of three boxes.

| Result | What it means | Priority |
| --- | --- | --- |
| **Accurate description** | Your entity is consolidated in the model’s data. | Work on the business queries, the ones where your name does not appear. |
| **“I don’t know”** | An honest answer, harmless to you. | A visibility project, nothing to repair. |
| **Another company, described with confidence** | The model has filled a gap. | Deal with it first. It is the only case that actively harms you. |

> **The automated measurement trap**  
> If you tool up this test, do not count an answer as a “citation” because it contains your name: the name is already in the question, and the model is just repeating it. A string search counts as cited an answer that says, word for word, “I don’t know this brand”. You need to check that the facts stated actually describe *your* organisation.

## How do you stop leaving a gap to fill?

Five workstreams. None requires an advertising budget, and all of them also help your traditional search rankings.

**Declare the entity, don’t just describe it.** [Organization](https://schema.org/Organization) markup that only carries a name and a sentence consolidates nothing. Add the head office address, the registration number, the VAT number, the founding date, the phone number. These can be cross-checked against public sources, and that cross-checking is what turns a string of characters into an organisation.

**Make authors verifiable.** An article signed with a first name, with no role, no employer and no photo, carries no authority. Role, affiliation, areas of expertise, a real portrait: the expertise you claim finally has someone to attach to.

**Corroborate both ways.** This is the point most often missed. Your site declares your profiles, and your profiles must point back to your site. A professional profile that links to a domain other than yours creates a contradiction where you wanted a confirmation. Check each profile, one by one.

**Remove ambiguities in plain words.** If your name looks like another brand or a game, if you have changed line of business: write it down. One explicit sentence cuts short the most likely confusion, and it is exactly the kind of sentence an engine picks up.

**Publish a machine-readable index.** An `llms.txt` file at the root, listing your reference content with its subject. It does not affect what the model has learned, but it serves the agents that read your site live.

## The market’s approaches, and who each one suits

| Approach | What it does | Who it suits |
| --- | --- | --- |
| **Do nothing** | Lets the engine improvise on every question. | Very established, extensively documented brands whose entity is already consolidated. |
| **Produce content** | Increases the material available, without guaranteeing it is attached to you. | Those who already have a clear entity and want to weigh on business queries. |
| **Technical markup** | Makes the entity declarable and cross-checkable. | Recent companies, ambiguous names, repositioning. |
| **Measure over time** | Turns an impression into a dated series, and lets you make decisions. | Those who want to know whether what they do has an effect. |

## The tooling

The manual diagnosis above is enough to know where you stand. What it does not give you is the series over time: the same test, on the same questions, at regular intervals, to see whether your corrections have an effect, and whether the engines themselves change, as they did between August and October.

That is what the generative visibility section of **SEO Hub**, published by Datafirefly Limited, does. The tool queries the engines on your queries, detects whether your brand is genuinely cited or just mentioned emptily, and keeps the history. It does not push your brand up: it measures. Method comes before the tool, and the requirements apply to any solution you might evaluate: does the measurement distinguish a citation from a mere repetition of the name, are readings dated and kept, and is the model queried recorded so that two measurements can be compared?

## What this says about the bigger picture

For twenty years, being visible meant appearing in a list of links. The reader clicked, landed on your site, and you were the one telling your story.

A growing share of searches now ends with a written answer. In that model, a third party summarises your story, based on what it thinks it knows, and that third party is not the same from one tool to the next: one will admit it does not know you, another will give you a different line of business. Your job changes in nature: give all of them accurate, structured, cross-checkable material, so that none has to improvise.

It is a new discipline, and it is entirely within your control: what your site declares, what your profiles confirm, and the consistency between the two. That is the work we do through our [artificial intelligence expertise](https://dotsland.com/en/expertise/artificial-intelligence/).

## FAQ

### What does an AI say when asked what it thinks of a company it does not know?

It depends on the engine. On 14 October 2026, asked about our agency without web access, GPT-4o and Claude Sonnet 5.5 said they did not know it in 16 answers out of 16, while Gemini 2.5 Flash described four invented companies in eight answers. None described the real agency.

### Can you get a wrong answer corrected?

No, and it would have no effect. The answer is not stored anywhere: it is generated differently for each question. What changes the result is what the model finds about you the next time it is trained or reads the web.

### Does this only affect small companies?

The risk is proportional to the lack of traces. An extensively documented brand is described correctly. A recent company, an ambiguous name or a recent repositioning are exposed. A name made of common words even more so.

### How long before it changes?

Nobody should give you a timeframe. Models are retrained on a schedule that is not disclosed, and they change from one version to the next: on our brand, Claude made things up in August and abstains in October. The only sound approach is to date your readings, record the model, and repeat the measurement at fixed intervals.

### Can Dotsland help?

Yes. We carry out the citability audit, consolidate the entity and set up measurement over time. Start with the four-question test above: it takes five minutes and tells you whether it is urgent. If the result worries you, [get in touch](https://dotsland.com/en/contact-us/).