The Retrieval Path — my method for AI search visibility

Why does a method need to be published at all?

Because you cannot inspect what you are buying otherwise. Answer engine optimisation is sold as a bundle of technical claims, most of which sound identical across proposals, and buyer guides now list a published, inspectable methodology as a screening criterion for exactly that reason. This page is the whole method, written in enough detail to be argued with on a call.

The Retrieval Path has five stages and they run in this order. The order is the load-bearing part: each stage is what makes the next one worth doing, and skipping one does not slow the others down so much as stop them compounding.

One scope note before the stages. The Retrieval Path is the method for the first of the three services on this site — getting found when buyers ask an AI. Entering a new market and turning interest into signed deals have their own sequences, set out on their own pages. This is not the method for those.

Stage 1 — Entity: make the name resolve to you

Before an engine can recommend a company it has to establish which company you are. That means one canonical description used identically everywhere the business appears, Organisation structured data with a stable identifier, and a complete and truthful set of profile links so the description is corroborated rather than merely asserted.

Where a name collides with another organisation — a similarly named company in a different sector, or a near-homophone of a large brand — this stage also covers disambiguation, because until it is resolved every later improvement strengthens somebody else's entity.

What you can inspect: the canonical description, the structured data, and the list of profiles with their live URLs.

Stage 2 — Prompt space: map the questions actually being asked

Keywords are not prompts. A buyer types "best flow meter suppliers in Europe for food-grade applications", not "flow meter supplier". So the second stage produces a written, versioned prompt set covering the shapes buyers actually use: shortlist requests, head-to-head comparisons, specification questions, procurement and compliance questions, and the "is X any good" question about you specifically.

The set is fixed before any work begins, and it is run once before anything on your site changes. That first run is the baseline every later run is measured against, and it is what makes a before-and-after possible at all. From then on the set is re-run unchanged: a prompt set edited between runs produces numbers that cannot be compared, and a prompt chosen after seeing the result is cherry-picking with extra steps.

What you can inspect: the full prompt list in writing before anything is measured, and the baseline run it produces.

Stage 3 — Structure: make passages retrievable

Retrieval systems work on passages, not pages. Content is split into chunks, embedded independently, and ranked against the query, which means each section has to survive being read entirely on its own. In practice that means every section opens with a direct answer, defines its own terms rather than relying on a definition three headings earlier, avoids pronouns pointing at earlier text, and sits under a heading phrased the way the question is asked.

The technical half of this stage is server-rendered HTML, structured data that matches the visible text, and a robots.txt with a named block for each AI user-agent — rules do not inherit from the wildcard block, so an agent named without rules is an agent with no rules.

What you can inspect: this website. Every structural rule described here is implemented on it, including the robots.txt and the structured data in the page source. It is the cheapest possible audit of whether the method is real.

Stage 4 — Corroboration: build the sources engines lean on

Your website is not the whole picture. Yext's October 2025 analysis of 6.8 million citations across 1.6 million AI answers found that 42% of citations came from listings and profiles against 44% from first-party sites. Roughly half of what an engine cites about you sits somewhere you may not have looked at in two years.

This stage covers the directories, industry listings, professional profiles and third-party pages relevant to your category — checked for accuracy, consistency with the canonical description, and completeness.

What you can inspect: the list of surfaces, with what each one currently says about you.

Stage 5 — Measurement: count citations, not impressions

This stage is re-measurement rather than the first measurement — the baseline was taken in Stage 2, before anything changed. The same prompt set is re-run on a schedule across ChatGPT, Perplexity, Google AI Overviews, Gemini and Claude, each recorded separately. Each run captures four things: whether you were named, whether one of your sources earned the citation, which source it was, and who was named instead of you.

The fourth is the one clients act on first. Knowing you were not named matters less than knowing who was, and what the pages behind those citations have that yours do not — the structure, the corroborating profiles, the specificity of the claim being quoted. Taking those apart is what turns a report into a work list.

Answer engines are non-deterministic. The same prompt can return different answers on different days and to different users, which is why the evidence is a trend across repeated runs and never a single screenshot. Anyone showing you one screenshot as proof is showing you one sample.

What you can inspect: the raw run data, not just the summary.

Why the order matters more than the tactics

Most of the individual tactics in this method are public. What decides whether they compound is sequence. Structuring passages on a site whose entity is ambiguous produces beautifully retrievable content attributed to the wrong company. Building corroboration before the canonical description exists spreads an inconsistent story faster. And changing the site before the prompt set exists leaves you with an after and no before.

That is also why this is one engagement rather than a menu. Answer engine optimisation, generative engine optimisation and AI search optimisation are the same five stages under the three names buyers search for, and SEO is the floor all of it stands on.

How long does this take, and where does one engagement end?

The audit runs about two weeks. The build that follows runs four to five. Those are the working shapes rather than a service-level commitment: a site with a rendering problem, or a company whose name collides with a larger brand, takes longer, and you hear that before the work starts rather than after.

The seam between the two sits inside Stage 3. Stages 1 and 2, plus the diagnostic half of Stage 3, are the audit — what your entity currently looks like, what the prompt set returns before anything changes, and which structural problems are in the way. The repair half of Stage 3, together with Stages 4 and 5, is the build. Stopping after the audit is a normal outcome, and the prompt set, the baseline and the findings are yours either way.

What I need from you

Less than most engagements ask for, but these four are load-bearing. The usual reason a four-week build takes seven is that one of them arrived late.

• CMS access with edit rights on the pages being changed — or a person who can apply changes within a couple of days.
• Admin on the Google Business Profile and the LinkedIn company page. Stage 4 is mostly work on surfaces you already own but may not have logged into in a while.
• One named point of contact who can answer a question without convening anyone.
• Thirty minutes with someone who actually sells. Stage 2 is only as good as the questions, and the marketing version of a buyer's question is reliably not the buyer's question.

The four questions to ask on a call

1. How exactly do you build the prompt set, and can I see it before you run anything?
2. What counts as a citation, and what happens when a company is named but not cited?
3. Which engines do you check separately, and what do you do when they disagree?
4. What does the report contain that I could not produce myself?

Those are the questions I would ask. Ask them of anyone you are considering, including me.