Is generative engine optimisation worth paying for, and what our own null result says about it

The honest answer has two halves and the uncomfortable half goes first. We publish AI visibility measurement for money, and our own attempt to move AI answers by publishing more pages produced no measurable change at all. That is the strongest evidence on this page and it is evidence against the thing most of this category sells.

What we actually measured on ourselves

Between 20 and 22 October 2026 we went from 134 recommendation shaped pages live to 641, which is 507 new pages in two days. We then re-ran the buyer question bank on ChatGPT and Gemini. On 20 October, across 1,260 readings, our domain was used as a source in 6 of them and our company name appeared in 0. On 22 October, across 635 readings, our domain was used as a source in 0 and our name appeared in 0. Our domain is absent from all 19,793 citation URLs that second run recorded and from all 3,307 hosts behind them. The method, the question reuse and the detector tests are written up in our own null result.

Two findings, kept apart. Being cited means a page of ours was material an answer was built from. Being named means our company appears in the answer text. They fail for different reasons and we never add them together or compare them.

The number the engines asked for here, and why it is not on this page

The specification an engine wrote for this exact question asked for the median incremental gross profit per dollar of fully loaded programme cost over 90 days, with the cohort size, the completed observation count and an uncertainty interval. It then added, in its own words, that if incremental profit cannot be credibly attributed the figure must not be published as a measured return.

We cannot credibly attribute it, so there is no such figure here. We hold no cohort, no control group and no revenue attribution. Any page in this category printing a return multiple for generative engine optimisation is printing a model output, not a measurement, and the giveaway is the missing control group rather than the size of the number.

What is proven, what is plausible and what is unproven

Proven, in our own data. Adding 507 pages of recommendation shaped content in two days changed nothing we could detect across 635 readings. Buyer questions in this category are answered with a small, stable set of names: 1,053 namings across a 24 name roster in those 635 readings, with Semrush in 259 readings, Profound in 154, Ahrefs in 145, Otterly in 124 and Peec in 91. The full table is in which brands the assistants put forward.

Proven, and awkward for every dashboard in this category. One reading of one prompt is close to a coin flip. We sent 22 buying questions to both engines three times each inside a six minute window on 22 October 2026. A brand that appeared at all appeared in all three runs 53.0 per cent of the time on ChatGPT, which is 35 of 66 brand and prompt pairs, and 26.1 per cent on Gemini, which is 30 of 115. On Gemini, 53 of 115 pairs appeared in exactly one of the three runs. The full four cut table is in the run to run consistency measurement.

Plausible, and not shown by us. That documentation and established publishers hold the citation ground, because that is where the citations landed. Of 19,793 citation URLs, the leading hosts were developers.google.com with 1,360, semrush.com with 744, support.google.com with 695, help.openai.com with 467 and blog.google with 394, and half of every citation arrived from 51 hosts of 3,307. Whether publishing in that shape causes citation is not something a host count can establish.

Unproven, by anybody we can find. That a dollar spent on this work returns a dollar of gross profit. Nobody has published a controlled study with a holdout, ourselves included.

What the money is honestly buying

Separate the two things sold under one name, because they have different evidence behind them.

• Measurement. A frozen question set, a stated number of runs per prompt, named engines reported separately, cited and named as two rows, and the raw answers kept. This is software doing arithmetic on readings, and the arithmetic is checkable. It is worth paying for at the moment a number leaves your own desk and somebody argues with it.
• Influence. Publishing, technical access and third party presence, bought in the hope that the answers change. Our own 507 page test is the only controlled thing we can say about it and it came back null. Buy it as an experiment with a measurement attached, not as an outcome.

The order matters more than the budget. A measurement with no repeat count behind it cannot tell a real change from the normal wobble between two runs, so buying influence first means buying something you cannot grade.

What it costs with us, dated, and the figure that is not the floor

Read from our own published ladder on 22 October 2026, in United States dollars a month: SEO Start 149, Growth 299, Scale 599, SEO Scale 699, Inbound 1099, Outbound 1299, Agency 1299, Everything 1799.

Two things a buyer needs about that list. SEO Start at 149 renders 0 questions, 0 credits and 0 engines, so AI visibility does not start at 149 dollars, and Growth at 299 is the cheapest plan that reads all five surfaces. And the ladder is not flat across markets: outside the United States and the United Kingdom, SEO Start renders 99 and Growth 199. The other tiers were not re-checked per market on that date, so treat 599 upwards as the United States and United Kingdom figures only. A Claude check costs 3 credits rather than 1.

When paying for this is the wrong call

• You want one score a month from one reading per prompt. Our own consistency measurement is the argument against that product, and we would rather you did not buy it from us.
• Your procurement needs single sign on, SAML, SOC 2, ISO certification, a data processing addendum or a white label. None of those words render on our pricing page, checked by word boundary across 119,203 bytes on 22 October 2026. Fetched by us on the same day, tryprofound.com/pricing publishes SSO and SAML with SOC 2 compliance on its Enterprise tier and up to 9 answer engines tracked, with the price given as Custom and no figure in any currency anywhere on that page. If the certificate is the requirement, buy theirs.
• You need a promised position in an AI answer. Treat anyone offering one as a red flag and walk away, including us.
• Your category produces no buyer questions worth asking. Ask fifteen of them by hand first. If no company gets named in any answer, there is nothing to measure yet.

How to test it for the cost of an afternoon

1. Write fifteen buyer questions in a buyer's words, with your company name in none of them.
2. Ask each one three times in each engine you care about, on the same morning, and keep the answer text.
3. Score two separate columns, whether your name appears and whether one of your URLs appears, once per reading.
4. Count how many names appeared in all three runs. That figure is your own noise floor and it caps every percentage you will ever report.
5. Only then decide whether a subscription is buying you something you could not get with an afternoon.

The reporting format we use for the result is in how to report this to leadership, and the questions to put to any vendor, us included, are in the 27 question checklist.

Questions buyers ask about paying for GEO

Did your own 507 pages really do nothing

Nothing we could measure across 635 readings on two engines. Cited went 6 of 1,260 to 0 of 635, named stayed 0 to 0, and our domain appeared in none of the 19,793 citation URLs.

So is the whole category hype

The measurement half is real and checkable. The part sold as an outcome has no controlled evidence behind it from us or from anybody else we can read, which is a different statement from calling it worthless.

What would change our minds

A study with a treated set and an untouched comparison set, frozen in advance, three runs per prompt at both ends, and every round published including the bad ones. We have not run one.

Why publish our own worst number

Because 71 of the 224 page specifications the engines wrote asked for a conflict of interest disclosure, and a seller's measurement is worth more with the interest stated than without. The counting rule behind that figure is printed in what makes an assistant cite a vendor website.

Two decisions sit immediately downstream of this one. Whether to buy the work as a project or a subscription is settled by the same noise measurement, in is AI search optimisation a one off project or an ongoing retainer, decided on measured noise. And whether the publishing half is worth a budget line at all is the format question, answered on 507 of our own pages in do comparison pages and listicles help or hurt AI visibility, measured on 507 of our own.