Clients who became the answer.

Two of the four case studies carry verified performance figures and two do not. That is a statement about evidence, not about the work.

Osoba Renos and Valve Appliance Repair both have Google Search Console data, so their pages quote movement from it directly — always as a percentage, always with the comparison window named beside it. The clients set that rule on their own data in September 2026: percentage movement is publishable, the click and impression volumes underneath it are not. Where a figure comes from Semrush rather than Search Console it is labelled an estimate, because a third-party tool modelling traffic and a Google property reporting it are not the same class of number and should never be printed as though they were.

YaDesign and NovelMaker carry no performance figures at all. Their headlines describe what changed structurally — one resolvable identity across two countries, a named category for a product that had none — because inventing a percentage for a named business is the single thing that would make every other number on this site worthless.

A renovation contractor, an appliance repair company, a bicoastal interior design studio and a consumer bedtime-story app have close to nothing in common commercially. The work ran in the same order regardless.

Retrieval first: whether the crawlers behind Google, Bing and the answer engines can reach and render the pages at all. Nothing downstream matters if they cannot, and this is where more engagements stall than anyone admits in public.

Then identity. Collapsing a business scattered across its website, a Google Business Profile and a dozen directories into one entity that resolves to the same facts wherever a machine looks. For YaDesign that meant two cities reading as one studio instead of two half-described ones. For Valve it meant a 5.0 rating, genuine parts and a same-day promise expressed as structured data rather than as a headline graphic no parser can read.

Then passages. Writing the answers to the questions people ask before they buy, in blocks that survive being lifted out of the page on their own. An answer engine rarely quotes a page; it quotes a paragraph, and a paragraph that needs the three above it to make sense does not get quoted.

Each case study carries an AI-visibility band listing what the answer engines can see, with the class of evidence marked against every line: a server log, a live page, or a prompt panel.

Only one of those three supports the claim clients most want to make. Server logs prove a crawler fetched the site — Osoba's record GPTBot, OAI-SearchBot, ClaudeBot and PerplexityBot doing exactly that. Structured data proves a machine can work out who the business is and that the details agree with each other. Neither proves an assistant recommends anyone.

That claim needs a fixed panel of prompts run on a schedule across every engine, with the result recorded each time, and it is measured inside an engagement rather than asserted on a marketing page. Where no panel has been run, the band says so in as many words. Blurring the line would be easy and a buyer comparing agencies would have no way to catch it, which is the reason it is worth not blurring.

Four is what LumiRank has taken to a standard worth publishing. A longer list would be easy to assemble and would mean less: portfolios pad with engagements that ended at a strategy deck, or with logos from work another team delivered.

Every client named here agreed to be named, links to a live site you can open and check, and is described in terms someone inside that business would recognise. Two of them let us publish their Search Console movement. Engagements that produce verified figures later will have them added, and the pages will not acquire numbers by any other route.