One vendor name turns up in all three of our readings, and no other does

Across three separate readings taken in this programme, one vendor hostname is present in every one. The host is semrush.com, and the second audit put it in 199 of its 1,260 readings. A differently written bank of questions, read the day before, put it in 95 of 597, split 84 on ChatGPT and 11 on Gemini. And in the separate ten city scan, its agency subdomain agencies.semrush.com was the most cited source in Los Angeles at 4 of 12 readings, and level at the top in Chicago and Mumbai at 3 each. Nothing else in our data does all three.

Persistence across three differently designed readings is a stronger signal than a high count in any one of them, and it is also the clearest evidence we have about what produces a count like that.

The three readings, and why agreement across them matters

Reading	Date	Size	What it found for this host
The first audit	15 October 2026	597 usable readings of 612 attempted, 102 questions per market	95 readings, being 84 on ChatGPT and 11 on Gemini
The second audit	16 October 2026	1,260 readings, all usable, 210 questions per market	199 readings, eighth in the top ten
The city scan	15 and 16 October 2026	120 readings, ten named cities, six buyer questions each	Top source in Los Angeles at 4 of 12, level first in Chicago and Mumbai at 3

The second audit's questions were written deliberately in eight directions the first one never went, including procurement depth, competitor named questions and comparison with adjacent search suites. A host that survives a question bank designed to go somewhere else is not there by accident of phrasing.

The engine split is the part most people will miss

84 against 11 in the first audit is the same vendor, the same day and the same questions, read by two apps. That ratio is not a property of the vendor at all. It is a property of the pairing between a vendor's pages and one engine's retrieval, and it means a single number describing how visible any company is across assistants has averaged two different results into one that describes neither.

• Report one column per engine. A combined figure hides exactly the information you would act on.
• Carry the denominator on each column. 84 of 306 and 11 of 291 are the honest pair.
• Expect the ratio to move. It is a reading of one date, and the point of writing it down is that the next reading is comparable.

Keeping that per engine record over time is the unglamorous half of competitor work, and it is the half that answers questions a one off screenshot cannot, which is the argument in keeping the log rather than the shortlist.

What a count like 199 is actually measuring

It is measuring how much of the category this company has published pages about, in a form something can quote. Four rows further down the same table sits another site of the same classification, searchengineinfo.org, at 147 readings, and very few buyers in this market would recognise it. Two rows of one type at wildly different levels of recognition means recognition is not what put either of them there, and published coverage is what is left.

That is a measured fact about a channel rather than a compliment to a competitor. We name competitors with their real numbers because a reader who sees us be accurate about somebody else has a reason to believe us about ourselves, and our own figure from the same 1,260 readings is 6.

What these three readings do not support

They say nothing about product quality, because nothing here is a test of any product. They hold no per industry rate, since the questions named no industry. The city rows are six questions per city and stop at the ten cities named in that scan. Two apps were read in both audits, ChatGPT and Gemini, while the product we sell reads five surfaces, so three of those surfaces are simply unmeasured in every figure above.

We sell AI visibility measurement and the publishing work that follows it, so the competitor counts and our own go on the same page. The Search Console side of our own record, with both windows and the export date, is our proof page.

Why the city rows are a subdomain and why that is not a quibble

The two audits cite the main hostname. The city scan cites an agency listing subdomain of the same company, which is a different set of pages doing a different job, and it is worth keeping the two apart in any table you build.

• The main host is reference and research. Dated figures, method pages, numbered findings, written for somebody deciding how the category works.
• The listing subdomain is a directory of suppliers. It answers who does this work in a named place, which is the exact shape of a local buying question.
• They win different questions. That is why one appears in the category audits and the other appears in a city scan, and why a table that merges them loses the distinction that explains both.

For a smaller company the second of those is the more interesting template. A directory page for one place, kept current, is a modest piece of work that answers a question a national brand page cannot answer, and in three of our ten cities a page of that shape was level with or ahead of everything else in the reading.

Questions about one vendor domain across three readings

How many readings cited semrush.com in the wider run?

199 of 1,260 on 16 October 2026. On the earlier and differently written question set of 15 October 2026 it was 95 of 597.

What was the engine split?

In the 15 October readings, 84 on ChatGPT out of its 306 readings and 11 on Gemini out of its 291. The two engines are reported separately because averaging them produces a number that describes neither.

Did the same company win any city in the city scan?

Its agency subdomain was the single most cited source in Los Angeles, at 4 citations of 12 readings, and was level at the top of Chicago and Mumbai at 3 each. Those are six question readings per city on both engines.

Does a high count mean the product is better?

No part of this is a product test. The count records how often a domain appeared in a source list, and the second vendor in the same top ten is a site most buyers have never heard of, which is the evidence against reading these rows as a quality ranking.

Can a smaller company reach a count like this?

Published coverage is what the rows reward, and the 147 reading row proves it does not require a known name. The scale of publishing behind a figure like 199 is years of it, so the honest version of the ambition is a count on your own questions that moves between one reading and the next.

Three readings, one name, and what to do with that

Pick the measurement you can repeat. A competitor appearing once in a screenshot tells you almost nothing; the same competitor appearing in three readings designed differently tells you where the published ground is. Write down which of your rivals survives your own second reading, because that short list is the real competitive set, and the rest was phrasing. Our first reading is set out in full, wording and controls included, at 597 readings across ChatGPT and Gemini in three markets.