Rebuilt a corporate site for LLM retrieval: cited in 6% of AI answers, then 74%

The brief: A European medical-device manufacturer noticed that clinicians and hospital procurement teams had started their research with an AI assistant rather than a search engine — and that the assistants were describing the company’s products from a competitor’s comparison page and a superseded product generation. Their own corporate site, which held the correct, regulatory-cleared specifications, was almost never the source. They asked us to find out why and fix it.

What the audit found

• We put the forty questions their sales team hears most to the major assistants and recorded every answer and citation. The company’s own domain was cited in 6% of answers. Competitor and third-party comparison sites carried the rest.
• Eleven of the forty answers contained a materially wrong specification — mostly figures from a device generation withdrawn two years earlier, still live on distributor pages the models trusted more than the manufacturer.
• The reason was mechanical, not editorial. The specification tables were rendered client-side and never appeared in the fetched HTML. The clinical evidence was locked in PDFs with no text layer. Nothing carried a date or an attribution, so nothing looked current.
• The site was, in other words, well designed for a human with a browser and unreadable to everything else. It had been optimised for a visitor who, as the comms team put it, had stopped coming years ago.

What we built

• A canonical product-facts layer. One server-rendered page per device holding every specification, indication, and clearance as plain text in a stable structure, each claim dated and attributed to the document that supports it. This became the single source the rest of the estate references rather than restates.
• Evidence out of PDFs. Clinical summaries and instructions for use republished as HTML alongside the original documents, so the substance is retrievable and the regulatory artefact stays intact.
• Structured data that matches reality. Product, organisation, and dataset markup wired to the same facts layer, so the structured claims cannot drift from the prose the way hand-maintained markup always does.
• A superseded-product protocol. Withdrawn generations kept online, clearly dated, explicitly marked as superseded and linked to the current device — rather than deleted, which had been quietly making the stale distributor copies the only surviving source.
• A retrieval monitor. A weekly harness that re-runs the forty questions across the assistants, records citations and flags any answer containing a specification that does not match the facts layer. The comms team gets a diff, not a dashboard.

What changed

• Own-domain citation rate across the forty tracked questions: 6% → 74% over five months.
• Answers containing a materially wrong specification: 11 of 40 → 1 of 40, and the remaining one is a third-party review site the monitor now flags for manual outreach.
• Organic sessions fell 22% over the same period. This was expected and was written into the success criteria before the work started — the assistant answers the question and the clinician does not click.
• Branded search volume rose 31% and demo requests attributed to “researched via AI assistant” in the qualification form went from a category that did not exist to 19% of inbound.
• Regulatory affairs signed off the facts layer as a controlled document. The claims a model repeats now come from the same source the submission does.

What we left behind

A corporate site that functions as the authoritative machine-readable record of what the company makes, and a weekly monitor that tells them when it stops being believed. The most valuable artefact turned out to be the least technical one: a single controlled place where a product claim lives, owned jointly by regulatory affairs and communications, which every other channel cites rather than paraphrases. The retrievability work is what made the models pick it up. The governance around it is what stops the problem recurring the next time a device generation is withdrawn.

We were candid with them about the limit of this work. Nothing here guarantees a citation — the retrieval behaviour of a model is not a surface anyone outside the lab controls, and a competitor doing the same work well will take share back. What the programme buys is the ability to detect within a week that a model is stating something wrong about a medical device, which for a manufacturer in a regulated market is the part that actually matters.