AI assistants were quoting two-year-old pricing and citing a directory

This is a worked example. It shows the measurement method and the reasoning we use on real projects, built from a pattern we encounter repeatedly. The figures are illustrative. We publish client numbers only when they are real, measured and the client has agreed to them being published.

Executive summary

A SaaS company had been asked by its board to report on "AI search". Nobody could define what would be reported, and two vendors had quoted for guaranteed placements — which is not a thing anyone can sell.

We built a measurement system instead: forty prompts across five types, frozen in wording, run across four AI search surfaces, recording six dimensions per answer. The baseline found the company mentioned in a third of answers, recommended in almost none of the ones that matter, and described using pricing it had abandoned two years earlier.

Business problem

The commercially important questions in this category are not brand questions. They are "what is the best tool for X" and "who should I use for Y" — asked by buyers who do not yet know the company exists. Those were the questions where the company was absent and a competitor was consistently named.

Separately, and more urgently: what the systems said about the company was partly wrong.

Baseline

The measurement system is the deliverable, so it is built first and then frozen:

• Discovery (8 prompts): "what are the best tools for [job to be done]"
• Category (8): "best [category] software for [segment]"
• Comparison (8): "[company] vs [competitor]", "[competitor] alternatives"
• Recommendation (8): "who should I use for [problem]"
• Brand (8): "what does [company] do", "is [company] any good"

Each answer was recorded on six dimensions: mentioned, recommended, cited, which sources were used, which competitors appeared, and whether anything said was factually wrong.

Freezing the wording is the part that makes this measurement rather than anecdote. These systems vary between sessions and change with model versions, so only rates over time on identical prompts mean anything.

Investigation

1. Run the baseline across every available surface. Record per prompt, per surface, on all six dimensions. One screenshot is not data.
2. Analyse which sources were used. In more than half the answers where the company appeared, the cited source was a third-party software directory — whose listing had not been updated in two years.
3. Audit entity clarity on the company's own site. Could a machine extract, without inference: what the company does, who it serves, which markets, pricing approach, integrations, and who works there? Four of those six required inference from marketing copy.
4. Check consistency off-site. Website, LinkedIn and three directory listings gave three different descriptions of the product's scope.
5. Check technical access. The pricing page rendered its tiers client-side from an API. In the raw HTML response, there were no prices at all.
6. Inventory citable material. The site had 90 blog posts and no original data, research or worked example. There was nothing distinctive to cite.

Evidence

Root cause

Not one cause. A stack of them, in order of how cheaply they could be fixed:

• The facts were not stated plainly anywhere. Positioning copy is not a fact. "Built for teams who move fast" does not tell a machine — or a buyer — who the product is for.
• The pricing page was invisible to anything that does not execute JavaScript. So the only available price was the two-year-old directory figure.
• Off-site descriptions contradicted each other, which makes every source about the company less usable.
• Nothing was worth citing. With no original material, the systems had to reach for third parties, and they did.

Actions

1. Publish a plain-language facts block on the About and each service page: what this is, who it is for, which markets, what it costs, what it integrates with. Acceptance: all six facts extractable from the raw HTML response with no inference.
2. Server-render the pricing tiers. Acceptance: curl of the pricing page returns the tier names and prices in the HTML body.
3. Correct the six factual errors at source — update the directory listings, LinkedIn and the profiles that were cited. Acceptance: each cited source now states the current facts.
4. Make the core description identical across every surface. One sentence, used verbatim on the site, LinkedIn and every directory. Acceptance: string match across all listed profiles.
5. Publish one piece of genuinely original material per quarter — data, research, or a worked example. Acceptance: something exists that a third party would have a reason to cite.
6. Verify crawler access in robots.txt for the relevant crawlers, and confirm the explaining pages render server-side. Acceptance: rendered-crawl diff shows no critical content missing from the initial response.
7. Re-run the frozen prompt set monthly and report the rate change.

Implementation

Items 1, 2, 3 and 4 are days of work, not months, and they carry most of the value — because they are also the ones that fix incorrect information reaching buyers right now. Item 5 is a standing commitment rather than a task.

Nothing in this list is an "AI trick". Google's own guidance treats AI search optimisation as part of SEO, and does not require AI-specific schema or an llms.txt file for its AI features. Every action above also improves conventional search and the experience of a human reading the page.

Result

What we measure, monthly, on the identical frozen prompt set:

• Mention rate across all prompts.
• Recommendation rate on category and recommendation prompts — the two types that reach buyers who do not know the brand.
• Citation rate with the company's own domain as the source.
• Competitor mention rate on the same prompts.
• Count of remaining factual errors — this is the one we expect to reach zero and stay there.

What we do not promise: that any system will mention or recommend the company. There is no lever for that, and a vendor selling one is selling something else. What we can promise is that you will know what is being said, that what is said will be accurate, and that the signals which can influence it are in good shape.

What we learned

• The factual-accuracy dimension is the most actionable finding, every time. It is fixable in days and it is actively costing something today.
• Client-side rendering of the facts that define you is a real problem. If your price only exists after JavaScript runs, the price that circulates is whoever else published one.
• Consistency across sources is a lever. Three contradictory descriptions make all three less usable.
• A single check is not a measurement. Freeze the prompts, run monthly, report rates. Anything else is anecdote with a screenshot attached.

What happens next

The measurement system is the durable asset. The prompt set, the competitor set and the monthly format outlive any individual model version — which matters, because the surfaces themselves will keep changing.