Case studies · AI visibility

Three stores, ninety days, and what the answer engines did next.

When a shopper asks ChatGPT, Claude, Gemini or Perplexity what to buy, the answer names a handful of brands. These three stores started outside that handful. Here is the baseline we measured, the work we did, and the trajectory we model over the first ninety days.

Boo, the LLMBOO ghost, looking through a magnifying glass
Read this first. Every day-0 figure is a real measurement from a live scan run in September 2026. Cases one and three are our own stores, which we run as testbeds for the engine described here. Case two is a third-party retailer measured from public data, not a client. The 30, 45 and 90-day visibility figures are modelled projections, not reported outcomes; the machine-readability figures are measured before and after. Store names are withheld.
The method

We work backwards from the answer.

Classic SEO optimises for a page in a list of links. An answer engine never shows that list. So we start at the other end: with what the engines actually say today, in your category, to a buyer who sounds like your buyer.

SEE

Ask the engines

We put your category's real buying questions to all four engines and record who gets named, in what order, and why.

DIAGNOSE

Read the reasons

We extract the attributes the winning answers lean on, then test your site for whether a machine can find those same facts about you.

FEED

Rebuild the machine layer

Our engine turns your catalogue into what assistants and shopping agents read: a plain-language store summary, a Markdown twin of every page, product feeds, and structured data with real prices and stock.

PROVE

Re-ask, and count

The same questions, on a schedule. The score moves or it doesn't, and you see which engine changed its mind.

The feed step is where our own system does the work a merchant cannot reasonably do by hand: it reads the catalogue, decides which facts an engine can quote and verify, and publishes them in the formats the engines fetch. The specifics of that mapping are ours.

Case one

A boutique audio-software studio

Software · EUDirect sales~20 products

A specialist studio selling audio plugins to producers. Strong word of mouth in forums, invisible to every assistant: across every buying question we put to all four engines, not one named it. The engines answered with the same three incumbents each time. The store already had a summary file and company data, but nothing an engine could quote about the products themselves.

AI visibility score, day 0 measured, then modelled
0 20 40 60 0 34 day 0 30 45 90
Day 0 measured. Days 30 to 90 modelled.
Engines naming the brand0 of 4→1 of 4
Machine-readability score60→95
Products an engine can quote0→all
Product structured datanone→live
Markdown pages for agentsnone→every page
  • Published a full catalogue in the plain-text format assistants read, with what each plugin does in a producer's words.
  • Added price, licence terms and system requirements as quotable facts, not marketing copy.
  • Wrote the comparison page the engines kept reaching for when naming rivals.
Case two

A speciality coffee roaster

Food & drink · UKSubscription~40 products

A well-reviewed roaster with a subscription business. One engine named it occasionally; the other three answered the same buying question with the same four competitors every time. Its site blocked nothing and had decent company data, but no machine-readable catalogue at all.

AI visibility score, day 0 measured, then modelled
0 20 40 60 23 52 day 0 30 45 90
Day 0 measured. Days 30 to 90 modelled.
Engines naming the brand1 of 4→3 of 4
Machine-readability score45→95
Rivals owning the answer4 of 4→2 of 4
Shopping-agent feednone→live
Roast & origin facts exposedimages only→text
  • Moved roast level, origin, tasting notes and grind options out of product images and into text a model can read.
  • Published a shopping-agent feed with live price and stock, in the formats the assistants ingest.
  • Answered the two questions the winning rivals answer and this roaster did not: subscription flexibility and freshness dating.
Case three

An own-brand apparel store

Apparel · EUOwn-brandOur testbed

The hardest category of the three: apparel, where assistants default to household names and marketplaces. We run this store ourselves precisely to test that case. For comparison, we measured a global sportswear brand's own site at 25 out of 100 for machine-readability. With our layer live, this store measures 80 and rising, while visibility remains the slow half of the equation.

AI visibility score, day 0 measured, then modelled
0 20 40 60 12 38 day 0 30 45 90
Day 0 measured. Days 30 to 90 modelled.
Engines naming the brand1 of 4→2 of 4
Machine-readability score25→80 today
Catalogue visible without scriptsno→yes
Sizing & materials as textpartial→full
Agentic checkout discoveryno→live
  • Republished the catalogue as plain text an assistant can read without executing the storefront.
  • Turned the size chart, fabric composition and care instructions into quotable facts per garment.
  • Targeted the narrow questions this label can actually win, rather than the ones the household names own.
How to read these numbers

What the score is, and what it is not.

The visibility score is the share of a category's buying answers in which a brand is named, weighted by how prominently and by which engine. A score of 100 would mean every engine names you first, every time, which no real brand achieves. In a competitive category, moving from single digits to the thirties means going from absent to routinely considered.

Machine-readability is measured separately: whether an engine is allowed to read the site, whether a catalogue exists in the formats they fetch, and whether prices, stock and product facts are stated where a machine can find them. It is the part we control directly, which is why it moves first and furthest.