In the 60 days to 25 August 2026 we asked ChatGPT, Google AI Overviews, Claude, Perplexity and Gemini 876 non-branded questions about the things we sell, and not one of those answers cited Argrow.
This post is the whole result, including the parts that do not flatter us. We sell AI-search visibility measurement, so the least we can do is run the measurement on ourselves and publish what came back.
What did we actually measure?
We run scheduled probes against five engines. A probe is one question, put to one engine, with the full response recorded so it can be re-read rather than remembered. Between 31 May 2026 and 25 August 2026 we collected 10,872 probes across the five, all of them from our own production system.
For this post we split our own probes into two buckets:
- Non-branded questions. The things a buyer types when they do not yet know we exist, like how to check whether an AI engine mentions their business.
- Branded questions. Questions that contain the word Argrow.
The split matters, because they answer two completely different things. One asks whether the engines will recommend us. The other only asks whether they have heard of us.
The uncomfortable result
Across the 60 days to 25 August 2026, the two buckets landed like this:
| Bucket | Probes | Times cited |
|---|---|---|
| Non-branded | 876 | 0 |
| Branded | 68 | 14 |
Zero. Not a low number, not a declining number. No engine reached for us once across 876 chances when the question did not contain our name.
The branded picture is different, and it is the reason we can say the pipeline works rather than that the probes are broken. Broken down by engine, our branded citations landed like this:
| Engine | Branded probes | Cited |
|---|---|---|
| Perplexity | 4 | 4 |
| Gemini | 3 | 3 |
| Claude | 4 | 3 |
| ChatGPT | 4 | 2 |
| Google AI Overviews | 53 | 2 |
Read the two tables together and the finding is plain: as at August 2026, the engines know who we are when asked by name, and never reach for us otherwise.
How can I track if my company appears in AI-generated answers?
The method is not clever, and you can run a rough version of it yourself. Write down the questions a customer would type before they knew your name. Put each one to each engine. Record whether your business is named, and keep the answer text. Repeat on a schedule so you are reading a trend rather than one roll of the dice.
Two things make this harder than it sounds. Answers vary between sessions, so a single ask proves very little. And the engines do not agree on what a question even deserves, which is the second thing our data showed.
The engines do not behave alike
Over the full 10,872-probe dataset from 31 May to 25 August 2026:
| Engine | Probes | Behaviour |
|---|---|---|
| Google AI Overviews | 7,494 | An AI Overview appeared on 3,400, or 45.4% |
| ChatGPT | 913 | Answered every probe |
| Perplexity | 911 | Answered every probe |
| Claude | 792 (since 20 June 2026) | Answered every probe |
| Gemini | 762 | Usable citation data on only 94; last probe 3 August 2026 |
The plain reading, as at August 2026: Google is the only one of the five that often declines to generate an AI answer at all. The others always answer something. That has a practical consequence for anyone building a report on this, because a Google-only view of AI search will show you an absence roughly half the time and you cannot tell from a summary number whether you lost the citation or the answer never existed. We go into what that means for measurement in the AI search visibility guide.
Gemini is the other honest caveat. We stopped getting usable citation data from it, our last probe was on 3 August 2026, and we would rather say that than quietly carry a stale engine in a chart. It is currently off in our own roster.
What we think it means
We do not think the engines have taken a view on us. We think we gave them nothing to quote.
We scored our own blog with the same content tooling we run for customers, and the result was consistent with the probe result. Across nine articles and roughly nine thousand words, our own writing contained 10 checkable data points in total, and seven of the nine articles contained none at all. Every article scored below our own 55-point E-E-A-T threshold, in a range of 22 to 47.
That is the shape of content an engine skims and drops. There was no price, no date, no measured figure, no first-hand result: nothing an answer engine could lift and attribute to us. We had written a lot of pages about being quotable without ever being quotable, and we did not notice until we pointed the measurement inwards. The mechanics of what makes a page liftable are in our academy guide on structuring a page so answer engines can quote it.
Google Search Console tells the same story from a different angle. In the full month of July 2026, argrow.ai took 311 impressions and 7 clicks. Our single largest query, “seo report dashboard roi”, drew 104 impressions, 0 clicks and an average position of 96.2. You cannot be cited by an engine that never surfaces you in the first place.
What we changed because of it
Four things, all of them already done as at 25 August 2026.
- We deleted three blog posts. Two competitor posts were cannibalising our SearchAtlas and OTTO comparison and, unlike that page, they were unsourced, undated and conceded nothing to the competitor. A third post duplicated the intent of our generative engine optimisation explainer. All of them now 301 to the page that should have had the traffic.
- We put a floor under our own writing. Every remaining article has to carry at least two checkable, dated facts, or it does not ship. That rule is enforced by the same linter our content agents run, so we cannot quietly exempt ourselves.
- We retired the head terms. As at 24 August 2026 we do not appear in the top 100 for “ai seo tool”, “ai seo platform”, “generative engine optimization”, “ai search visibility”, “seo platform” or “seo automation software”. Those queries belong to companies with a decade of head start and a budget to match. We retargeted our pages at the long-tail questions Search Console shows we already touch, like the one this post is written for.
- We started dating facts inside the sentence that carries them, rather than only in a byline, so a reader and an engine can both see how old a number is without hunting for the page footer.
None of that is a strategy we invented for this post. It is the same work we would put in front of a customer looking at a flat AI-visibility chart, applied to ourselves, and we will report on whether it worked using the same baseline-adjusted method we use for everyone else.
What we will publish next, and when we re-measure
We are leaving the probe schedule exactly as it is. Changing the questions at the same time as changing the pages would make the next number unreadable.
We will re-run the same 876-question non-branded set and publish the result 90 days from this post, in late November 2026, whatever it says. If the number is still zero, that will be here too, along with what we got wrong. The next post in this thread will cover which specific pages, if any, earned a first citation and what those pages had that the others did not.
If you want the honest version of this measurement pointed at your own site rather than ours, that is what AI search visibility tracking does. It will tell you the same kind of thing it told us, and you may not enjoy the first report either.
Frequently asked questions
How can I track if my company appears in AI-generated answers?
Put the questions your customers would actually type to each engine on a schedule, record whether your name appears in the answer, and keep the raw responses so the result can be re-read later. A single check is a snapshot; the value comes from repeating the same questions so you can see the number move. That is exactly the method we used on ourselves in the 60 days to 25 August 2026.
Does being cited for your brand name mean you have AI visibility?
No, and our own data is the clearest example we have. In the same period our branded questions were cited 14 times out of 68 probes, while our non-branded questions were cited 0 times out of 876. Being recognised when you are named is a much lower bar than being recommended when you are not.
Why does Google AI Overviews behave so differently from ChatGPT?
Because Google frequently declines to generate an AI answer at all. Across 7,494 Google AI Overviews probes between 31 May and 25 August 2026, an AI Overview appeared on 3,400 of them, or 45.4%. ChatGPT, Perplexity and Claude answered every probe we sent them in the same window.
Is zero citations a normal result for a new site?
We cannot tell you what is normal, because we only have our own numbers and we will not publish anyone else's. What we can say is that it is a plausible result for a business whose pages carry very little that an engine can verify and quote, which is what our own content audit found.