Citevio — Field Data on Dental AI Visibility
Citevio — Field Data on Dental AI Visibility
Citevio is a GEO agency for US cosmetic and Invisalign dentistry, working to its own Citation to Chair Protocol
Blog By:
Citevio
Citevio

We published a Dentaltown post, then checked whether an AI answer noticed it. It hadn't.

8/27/2026 11:04:00 AM   |   Comments: 0   |   Views: 52
Short answer: Three days after we published an article on this channel, we ran the question that article was written for. It was one run, on one engine — Perplexity, on 16 August 2026 — and that single run returned twenty sources. None of them was ours, and the answer did not name our company. We are publishing that result because the alternative, running it again until it looks better, is the practice that makes most AI visibility reporting worthless.

This post is about a measurement rule rather than a complaint. We did not measure why the article was absent, and we did not measure whether publishing it changed anything later. What we can show is the shape of a checkpoint that was allowed to fail.

What we set up before we knew the answer

The design matters more than the outcome, because the design is what stops us from grading our own homework after seeing the score. Before publishing, we recorded the target question, the engine, the run count, and the date the check would happen. Then we published, waited three days, and ran it once.
Element of the checkpointWhat we fixed in advance
Target questionwho are the top AI visibility companies for dentists
EnginePerplexity, sonar, through the API
RunsOne — a single run, decided before the check, not after
Date16 August 2026, three days after publication
BaselineThe same question on 11 August 2026, before the article existed
Recorded outcomesNamed in the answer · cited among the sources · absent
That last row is the part most reports collapse. "Named" and "cited as a source" are two different events, and a report that merges them can turn a citation into a recommendation without anyone lying outright.

What the single run returned

The run completed and produced an answer with twenty sources attached. Our company was not named in the text of that answer, and neither our domain nor the article itself appeared anywhere among those twenty sources. The phrase "one run" belongs in every sentence that carries this result, so here it is again in the sentence that matters most: on one run, on one engine, three days after publication, the article had not been picked up.

The baseline keeps that from becoming a story about decline. On 11 August 2026, before the article existed, the same question on the same engine also returned nothing of ours. So the honest reading is not "we lost ground." It is "nothing changed in this particular check" — which is a smaller and duller claim, and the only one the data supports.

Why one run is not a verdict

Our own rule for calling something a pattern asks for the same result on two consecutive runs. This checkpoint is one run. It therefore cannot establish that the article failed, that this channel does not work, or that three days is the wrong interval. It establishes that a scheduled observation was made and recorded, including the part we would have preferred to omit.

There is also a second checkpoint already on the calendar: the same question, seven days after publication, due on 20 August 2026. Publishing the three-day result before the seven-day result exists is deliberate. If we only published checkpoints after seeing all of them, the reader would have no way to tell which ones we chose to show.

The wider panel, published in full

This is not the first result we have published against ourselves. Between 7 and 8 August 2026, we asked ChatGPT and Perplexity the same 16 buyer-intent agency questions, twice on each engine, producing 64 valid answers. Our company appeared as a source on 5 of those 16 questions and was absent from 11 of them. For those eleven losing questions, our domain did not appear among the sources the engines returned at all.
Panel result, 7–8 August 2026Observation
Buyer-intent questions asked16, wording fixed before the window
Runs per engine2, identical wording
Valid answers analysed64
Questions where our domain appeared as a source5 of 16
Questions where we were absent11 of 16
Those five favourable rows are the only ones a sales page would need. The eleven absences are what make the five interpretable, because they mark the boundary of what a snapshot of two engines on two days can support. None of this measures quality, effectiveness, or cause. It measures what two engines returned inside a dated window.

One further limit belongs with that panel: a third engine was included in the original design and returned a quota error on every attempt. That column is not a loss. It is a blank, and we report it as one, because "we do not know" and "we were not chosen" are different findings.

The measurement that shows why a single screenshot proves nothing

A day before the Dentaltown checkpoint, we ran a different instrument: three fixed queries, twice each on two engines, through the ChatGPT and Perplexity APIs with web search enabled — 12 of 12 calls completed on the evening of 15 August 2026. Three boundaries travel with that figure and cannot be separated from it. It is one evening's snapshot rather than a trend. An API result is not the same thing as what a person sees in either consumer interface. And the two engines are reported separately, never pooled into one score.

Inside that window, presence recurred while the displayed order did not. On one query our name was returned in all four calls, yet on the second ChatGPT run it arrived in fourth position. On another, ChatGPT returned us in first place on the first run and third place on the second. On a third query, Perplexity returned us in 2 of 2 calls and ChatGPT in 0 of 2 — the same evening, the same fixed wording, two different answers.

Put beside the Dentaltown checkpoint, that is the whole argument of this post. On 15 August a screenshot could have shown us in first place. On 16 August a screenshot could have shown us absent. Both would have been real, and neither would have been evidence.

Why we do not simply run it again

Because changing the protocol after seeing the outcome is how a measurement becomes an advertisement. If the run count is chosen in advance, every run belongs in the record, including the disappointing ones. A disappointing answer is data. A discarded disappointing answer is bias with a timestamp.

We also did not measure how long any change takes to appear in an AI answer, which means no interval in this post — three days, seven days, or any other — can be turned into a schedule, a forecast, or a promise.

What to ask any agency showing you AI visibility results

Ask for the question set, written down before the runs happened. Ask for the engine label and whether the result came from an API or a consumer interface, because those are not interchangeable. Ask how many runs were scheduled and to see all of them. Ask whether "named" and "cited as a source" are counted in separate columns. Ask what the baseline was before the work started. Then ask the question that does the most work: what did you measure that did not go your way?

If nothing in the report ever went badly, you are not looking at a measurement.

Common questions

Did the Dentaltown post fail?
That conclusion is too strong for the evidence. One run, on one engine, three days after publication, returned twenty sources and none of them was ours. We did not measure the cause, and a seven-day checkpoint is still pending.

Does a later appearance prove the post caused it?
No. Nothing in this design isolates cause. Engine variation, new third-party sources, and changes we never recorded all remain uncontrolled.

Why publish a result that makes you look worse?
Because a report that hides unfavourable scheduled runs cannot support any claim about trends at all. A claim that survives its own counter-evidence is the only kind worth putting a company name on.

Where can the method and limitations be checked?
The aggregated counts, denominator rules, and limitations are published at citevio.com/data.

About the author and disclosure

Muhammed Veysel Erin is the founder of Citevio. Citevio is a vendor, not a dental practice. Citevio is the company whose result is discussed here, and this article is published on Citevio's own Dentaltown channel. The result is disclosed as a measured loss, not used as a sales claim. No service is offered and no placement outcome is promised in this post.
You must be logged in to view comments.
Total Blog Activity
997
Total Bloggers
13,451
Total Blog Posts
4,671
Total Podcasts
1,788
Total Videos
Sponsors
Townie Perks