What generative engine optimisation work actually consists of, from our own run

The engine that specified this page wrote down why it exists: agency pages in this category describe strategy roadmaps and content optimisation without itemising the operational workflow. It then offered its own headline figure, that standard enterprise retainers commit 40 hours a month across four pillars. We do not have that figure, nobody handed us the engagements behind it, and it is not repeated here as a fact.

What we can publish instead is one real run of this work, counted. It is our own, it was finished on 22 October 2026, and the parts that went wrong are in it.

The run, in numbers

• Questions. The bank holds 324 written questions. 318 were asked. The remaining 6 are brand questions that name our own company, and that batch was written and never run, so no prompted figure exists anywhere in this run.
• Batches. 39 batch identifiers on the measurement arm, and 19 batches on the in conversation pipeline arm.
• Readings. 635 scored. 550 of them ran a live web search and the rest answered from training data.
• Citations collected. 19,793 URLs, every one with a parseable host, across 3,307 distinct hosts. A further 2,583 citations came from the pipeline arm.
• Engines. ChatGPT and Gemini on the measurement arm. Claude appears only on the specification arm and only in the product interface, which produced 6 of the 224 specifications.
• Specifications produced. 224. By arm, 37 from the full in conversation pipeline and 187 through the interface. By engine, 104 from Gemini and 83 from ChatGPT through the interface, 31 from ChatGPT in conversation and 6 from Claude in conversation. 162 arrived with a full outline and 202 stated the dataset the page must contain.

The five operations, and which one cannot be automated

One, write the questions and freeze them

Questions in a buyer's words, with the company name in none of them except a final batch that is ordered last. In this run no question in batches 1 to 18 names our domain. This is a writing job, it is where the measurement is won or lost, and it is the cheapest thing to get wrong: a question set assembled after seeing results measures the person who assembled it.

Two, lock the market in the prompt text

This one is a finding rather than a task. The interface we read the engines through accepts four location parameters, gl, hl, location and country, and ignores all four. We measured that. So a United States reading is produced by a market lock line inside the prompt itself, not by a setting, and any vendor telling you their geography comes from an API parameter is describing a field that had no effect in our hands.

Three, ask and record

Each question sent as a single turn with a live search requested, the answer text kept, the sources block kept separately from the prose. A name under a Sources heading is a citation. A name in the prose is a recommendation. Counting them together is the most common way a report in this category becomes unreadable.

Four, audit the engine's own answer, by hand

This is the step that will not automate, and the reason is a character count. The audit step feeds the engine its own answer back and asks which sources it used. The interface caps a prompt at 4,000 characters. One batch answer in this run measured 21,711 characters. The step therefore ran by hand, in the products, by the founder, for 38 pipelines. If a vendor tells you their pipeline is fully automated end to end, this is the step to ask about.

Five, score with a written rule

Once per reading in which a name appears, never once per occurrence. That rule choice changes the totals, so it is published rather than assumed. On this run the result was 1,053 namings of rivals across a 24 name roster, 451 readings in which a rival was named and we were not, and 184 readings in which no company was named at all.

What the run produced, including the part that failed

Our own company was named in 0 of the 635 readings and our domain appeared in 0 of the 19,793 citation URLs. Two days earlier, on a larger bank of 1,260 readings, we had been cited in 6 and named in 0, with 134 recommendation pages live. By the 22nd we had 641 live. So the publishing half of this work, 507 pages in two days, produced no measurable change, and that is written up in the 507 page null result.

The honest conclusion from one run of the whole operation is that the measurement half is real work with checkable output, and the publishing half is an experiment whose result in our case was zero.

How to turn this into a scope of work you can sign

1. Annexe one, the question set, with its count, frozen and dated before anything else starts.
2. Runs per prompt, written as a number. Three per prompt per engine is the minimum our own consistency work supports, measured in the run to run measurement.
3. Engines and interface named. An answer through an API with no live search is a different object from the app your buyer uses, and in this run 85 of 635 readings had no live search behind them.
4. Two outcome rows, cited and named, never merged.
5. Raw answers retained and exportable. The text is what tells you what to fix. A stored percentage does not.
6. Who does the step that cannot be automated, and how many hours of it are in the price.

What this does not cover

One programme, our own, finished on 22 October 2026, two engines on the measurement arm. No hours log, so there is no labour figure here and no cost per reading. No comparison set. We sell this work, so read the operational detail as a seller describing its own process and check the parts that are checkable, which are the character cap, the ignored location parameters and the published prompt set.

On our own gaps: we read five surfaces, and our pricing page renders no SSO, SAML, SOC 2, ISO, DPA or white label, checked by word boundary over 119,203 bytes on 22 October 2026. Fetched the same day, tryprofound.com/pricing publishes up to 9 answer engines tracked on Enterprise with SSO, SAML and SOC 2, and no price in any currency.

Questions about how this work is actually done

How long does one measurement take

It is a multiplication rather than a duration. Questions times engines times runs. This run asked 318 questions and produced 635 scored readings across two engines.

Which part is software and which part is a person

Asking, recording and scoring are software. Writing the questions is a person. Auditing the engine on its own answer was a person in this run, because one answer was 21,711 characters against a 4,000 character prompt cap.

Why were 6 questions never asked

They are the brand questions, which name our own company. The runner skips every brand row in both its validation and its job loop, so the prompted ceiling was never measured and no prompted figure appears anywhere in this run.

Can I repeat this myself

The full question set from the measurement arm is published with the citation host study, so the run can be repeated against your own domain without asking us for anything.

Two parts of this run are worth reading on their own. The measurement limits that shaped it are in which AI surfaces can actually be measured through an API, and which cannot, and the dataset hygiene decisions behind its citation table are in two of the twenty most cited hosts in our citation data were not sources at all.