Full protocol, ten anonymous sessions, 1 August 2026
We asked ChatGPT ten buying questions about our own market
Most companies check what an assistant says when a buyer asks for what they sell, and then never mention the result, especially when it is uncomfortable. We ran the check on our own category, wrote down every answer, and are publishing the whole protocol, including the nine questions where our own site was not named once.
· 7 sources
How this was measured
Ten buying questions, asked on 1 August 2026. Each one in a fresh session with no history, so the assistant could not lean on anything discussed earlier. None of the first nine contains our name, our brand or any of our products: a question that contains your name tests recognition, not visibility. The tenth does contain the name and exists purely as a control, to show the difference between the two. Every answer was recorded in full, along with every company named.
| Buying question | Named us | Who it named instead |
|---|---|---|
| Who can build a working MVP in a week in Europe? | no | Pixels & Logic, Minimum Code, Design Sprint, 10Clouds, bluesBrackets, Altar.io |
| Is there a good AI expert or AI agency in Slovenia? | no | Abelium, Pareto AI, Valira AI, Jožef Stefan Institute AI Lab, SLAIS, SLOAI |
| How much does an AI voice agent cost in Europe? | no | Vapi, Retell, Bland, Synthflow, Twilio, OpenAI Realtime |
| Who does GEO, getting a product recommended by ChatGPT? | no | First Page Sage, Position Digital, Go Fish Digital, Siege Media, Single Grain, iPullRank, Omniscient Digital, Scandiweb, Profound, Evertune, Peec AI, Scrunch AI, Otterly AI |
| What tool turns property photos into a sales video? | no | PhotoAIVideo, Amplifiles, AutoReel, Revid |
| What does a fractional Chief AI Officer cost per month? | no | Whitesmith, Jose Miguel Sanchiz, Boardroom AI, The Professor-AI, Iternal |
| What does a short-form video content factory cost? | no | Descript, ElevenLabs, Adobe, Epidemic Sound, Upwork |
| Does llms.txt matter, and what actually works? | no | llmstxt.org, Answer.AI, OpenAI docs, Perplexity docs |
| Where are published AI price benchmarks for DACH and CEE? | no | ISG Provider Lens, Gartner, PAC RADAR, Clutch, GoodFirms, DesignRush, Malt, Hays, Eurostat, TED |
| Who is Mykhailo Koblychenko? (control question) | yes, first | mykhailo.koblychenko.com, LinkedIn, Bizi.si, CompanyWall.si, Luma |
The result: one out of ten, and the one was the control
On nine buying questions the assistant did not name us once. On the tenth, the one containing the name, our site was cited first and described almost verbatim from its own meta description. That gap is the entire finding, and it is sharper than a general complaint about needing more content: the entity is known, machine-readable and correctly extracted. It is simply not attached to the categories in which people ask.
This is worth separating from both SEO and brand awareness. Being known is not being recommended. A site can be perfectly readable, perfectly marked up and perfectly indexed, and still be absent from every answer that matters commercially.
Four things the answers had in common
- It cites primary price pages, not roundups. On every cost question the assistant linked vendor pricing pages, official crawler documentation, a central-bank exchange rate. Not one agency blog post summarising the market. If your prices are not published, you are not in that answer.
- It prices senior time in days. For fractional roles it built every band around days per month, and a monthly figure with no stated load is not comparable at all. An unscoped price is not a high price, it is an unreadable one.
- It hedges, and it hedges honestly. The answer about getting recommended by ChatGPT ended with an explicit caveat that nobody can make an assistant recommend a product on demand. Any offer promising the opposite now contradicts the very assistant the buyer is likely to consult.
- For a solo expert it defers to directories. Agencies were named individually. For an independent specialist it pointed at professional communities and listings instead. That is a structural difference, and it decides where a one-person practice has to be present.
What we changed because of it
- Stated the load on the fractional role. The assistant compares by days, and we had published a monthly figure with no scope attached.
- Explained the unit economics of the content service. The market reference is a per-video price an order of magnitude above ours, and an unexplained gap reads as low quality rather than as a different production model.
- Added dates and conditions to our own proof. An undated claim in the present tense eventually disagrees with a reader who checks it.
- Moved third-party presence to the top of the backlog. Three unrelated questions independently pointed at the same gap: directories and editorial mentions. Everything else on the list was already done.
What this measurement is not. One assistant, one date, one language, ten questions, one anonymous session each. Assistants answer differently depending on phrasing, language, model version and whether they search the web at all. A single run is a baseline, not a trend, and it is published as one. The same ten questions will be repeated on a fixed schedule across four assistants, and the protocol stays in the open, including the parts that do not flatter us.