We published 507 more pages, measured whether ChatGPT and Gemini changed what they say about us, and they did not

Between 2026-10-20 and 2026-10-22 we published 507 more recommendation shaped pages, taking the live count from 134 to 641, and then re-ran the same measurement on the same question banks. Before: cited in 6 of 1,260 readings, named in 0. After: cited in 0 of 635 readings, named in 0. The 507 pages moved neither number in our favour.

We are a company that charges money to improve exactly this. Publishing the result anyway is deliberate, and the rest of this page is the method, the verification, and the four things this result does not entitle us to say.

The two passes, side by side

Measurement	Pass 2, 2026-10-20	Pass 3, 2026-10-22
Recommendation shaped pages live	134	641
Readings scored	1,260	635
Readings that CITED our domain	6	0
Readings that NAMED us in the answer text	0	0
Rival namings recorded	1,314	1,053 across a roster of 24
Markets	United States, United Kingdom, India	United States only
Engines	ChatGPT and Gemini	ChatGPT and Gemini

Two columns of that table are the same measurement and two are not, so read it carefully. The cited and named rows are the same two detectors on the same two engines. The readings row is a different denominator, because pass 3 is one market where pass 2 was three. The rates are therefore not comparable as rates, which is why the counts are printed and no percentage appears in the table at all.

What is comparable is the question set. The evidence file records the pass 2 bank as 72 core and 210 US questions reused verbatim in pass 3, so 72 core questions and 210 United States questions came back word for word in pass 3. That is what makes this a before and after on one bank rather than two separate studies.

Cited and named are two findings and they stay apart

Cited means a URL on our domain was in the sources an engine returned. Named means our company name was in the prose a reader sees, with the sources block removed before scoring so that a name under a Sources heading counts as a citation rather than as a recommendation. The two have different causes and different fixes.

Pass 2 produced cited 6 and named 0. Collapsing those into one visibility figure would have hidden the whole problem, because a vendor can be cited as a source while no buyer ever hears its name. We have published the same separation on an earlier dataset in what 1,260 readings show about which names get repeated.

How the zero was verified before it was believed

A zero is the easiest number in the world to produce by accident, so it was checked three ways before it went on a page.

• The detectors were tested against planted text first. A detector that finds nothing is indistinguishable from a detector that is broken, so both were given answers with our name and our URL inserted, and both found them.
• The citation list was searched whole. Our domain appears in 0 of the 19,793 citation URLs this pass recorded, and it is absent from all 3,307 distinct hosts those URLs point at.
• The answer text was searched separately from the citations. 0 of 635 readings carry the name.

The same pass recorded 1,053 rival namings across a roster of 24, counted once per reading, so the detectors were plainly finding names. They were finding other people's, and the names they found are published with their counts.

What this result does not entitle us to say

Four claims a reader could reasonably pull out of the table, and why none of them is supported.

• It does not say content work has no effect. It says 507 pages of one shape, published over two days, did not change two counts on two engines inside that window. Two days is a short window for an index.
• It does not say 0 of 635 is worse than 6 of 1,260. The denominators are different and one pass covered three markets. Both are low counts and the honest summary is that both rounds found us close to invisible on our own measurement.
• It does not describe five surfaces. Both passes read two engines. Our product reads 5, so Perplexity, Claude and Google AI Overviews are outside this result entirely.
• It does not say the pages were wasted. Citation in an assistant and clicks from a search engine are different measurements on different surfaces, and the second one we publish from a Search Console export on our proof page rather than from this audit.

The three readings of this result that are worth having

Each of these is a hypothesis, and each one is testable by somebody other than us.

Index lag. The pages went live within days of the measurement. A page an engine has not retrieved cannot be a source, and 550 of 635 readings in this pass ran a live search at all. The next pass on the same bank is the test, not an argument.

The shape of what gets cited. In this same pass, half of all 19,793 citations came from 51 hosts out of 3,307, led by platform documentation and large tool vendors, and the whole host table is published in which websites these two engines cite for an American buyer. A new page on a new domain is competing for the half of the dataset where single citations live, which is a different job from publishing more of them.

Being named may not come from your own pages at all. 451 readings named a rival and not us, and 184 named nobody. The readings that named nobody are the openings, because nothing has to be displaced for a name to appear in them.

Why a vendor publishes its own flat result

Almost every number in this category is published by somebody selling the fix, including every number on this page. The only thing that separates a measurement from a brochure is whether the measurement is allowed to come out badly, and ours did, twice now. The earlier one is our own domain was cited 3 times in 597 readings, and it was published for the same reason.

It also sets a floor nobody can quietly restate later. 0 of 635 on 2026-10-22, on two named engines, against a frozen question bank with 72 core and 210 United States questions reused word for word. The next pass is measured against that, including by anybody who repeats it from the published question set.

What we changed because of it

The plan that came out of this pass is not another tranche of the same shape. The engines in this run were also asked what a page would have to publish before they would cite it, which produced 224 page specifications, 202 of them carrying an explicit dataset requirement and 162 carrying a full outline. Not one asked for more opinion. They asked for a stated sample size, a date, a disclosed method and a denominator, which is what this page carries and what the 507 pages mostly did not.

What this measurement does not cover

• Two engines on both passes. ChatGPT and Gemini. Five surfaces are in the product, two are in this result.
• Two dates, two days apart. 2026-10-20 and 2026-10-22.
• Different denominators. 1,260 readings across three markets, then 635 in one. Counts are comparable, rates are not.
• One reading per question per engine in each pass. The drift between two identical prompts on one morning is measured separately, in how often two identical prompts name the same brands, rather than assumed to be zero.
• The page count is our own inventory. 134 and 641 are counted by slug or by heading on our own site, which is a count of what we published rather than of what any engine retrieved.
• We sell the remedy. This page measures a method we charge for and reports that it did not move two counts in two days. Check the method rather than the conclusion.

Who should not buy the product behind this page

If what you want is a supplier who will tell you that publishing pages gets you named by an assistant inside a fortnight, we have just published the measurement that says otherwise, and you should buy from somebody else. If your procurement requires single sign on, SAML, SOC 2, ISO certification, a data processing addendum or a white label, our pricing page renders none of those words by word boundary on 2026-10-22, and Profound's own pricing page read on 2026-10-22 at tryprofound.com/pricing publishes SSO, SAML and SOC 2 on its Enterprise tier alongside up to 9 answer engines tracked.

And if you are shopping on price for AI answer monitoring, read the plan rather than the headline figure. SEO Start at $149 renders 0 questions, 0 credits, 0 engines. Growth at $299 is the cheapest plan that reads all 5 surfaces.

Questions about this before and after

Is 507 pages in two days not simply too fast to measure

Probably, and that is the single most likely explanation on the table. It is also the reason the pass is published rather than held: a dated flat result with the window stated is what makes the next pass readable as evidence instead of as a claim.

Why not wait and publish only the pass that works

Because a benchmark that only appears when it flatters its author is worth nothing to the reader, and every engine specification this run produced asked for failures to be shown beside wins.

Were the pages any good

They passed a gate that refuses invented figures, promised results, long dashes and known boilerplate, and each one carries its own pooled queries so that two of our pages never chase one question. What they mostly did not carry is original measurement with a stated N and a denominator, which is what the engines in the same run asked for.

Does a count of zero mean the detector failed

It was the first thing checked. Both detectors were given planted text containing our name and our URL and both found them, and the same detectors recorded 1,053 rival namings in the same pass.

What would change this finding

A third pass on the same frozen bank, with the same two engines and the same two detectors, after the pages have had time to be retrieved. That run is comparable to this one by construction, which is the point of freezing the bank.