We Read 3,553 Pages From 20 AI Search Agencies. 14 Name a Client Whose Leads or Pipeline Came From AI Search.

The one test

A mention is an appearance. Your brand shows up when a buyer asks ChatGPT for the best tool in your category. Visibility counts how often that happens. Share of voice counts your slice of the brands named. Citations count the URLs the engine quoted.

A commercial event is different. Someone books a call, requests a demo, signs up for the product, becomes a qualified lead, enters pipeline, or pays. It has a date and a person behind it.

So the test is one sentence: does the agency's own site carry a number on a commercial event, for a named client, credited to AI search specifically?

It is applied in that order. Is there a commercial number, is the client named, and only then, what is it credited to. A figure whose client cannot be looked up never reaches the first column, however it is attributed.

AI search means a named engine or surface: ChatGPT, Perplexity, Gemini, Claude, Google AI Overviews, or a phrase like AI search, LLM search or AI discovery. Four verdicts come out of it.

• AI-search pipeline. All three parts hold. A named client, a commercial number, and AI search named as the source.
• On the line. The number exists and the client is named, but the figure does not separate AI from organic search: either it covers both at once, or it carries no channel of its own while AI and organic figures sit beside it.
• Organic only. A commercial number is published and credited to organic search, or no client is named for it. Every agency here can show a lead or revenue figure somewhere. The question is what carries it and who it belongs to.
• Visibility only. No commercial number anywhere in the corpus. On this set, nobody.

Read three pages of an agency's site and plenty look like they only report appearances. Read the whole results set and all twenty have a commercial number somewhere. What separates them is what the number is credited to, and whether anyone is named.

Every agency, in alphabetical order

How the corpus was built

The method is a script. Five earlier hand passes over these same agencies each missed a page a later pass found: a home page, then a case-study index, then individual case studies behind an index. Every miss moved a verdict. So the enumeration stopped being a human job. For each agency, audit-agency-corpus.mjs reads robots.txt, walks every sitemap and sitemap index it declares, collects every HTML URL on that domain, and fetches up to two hundred of them once each, highest priority first, recording the HTTP status, the byte count and a sha256 of the response. The manifest, carrying a hash for every page, is the evidence behind this audit and we will send it to anyone who asks for it.

What the numbers mean. 16,528 URLs discovered, 3,701 fetched, 3,553 readable HTML pages, across 20 agencies. Six sites were small enough to fetch every discovered URL: Graphite, LoudFace, Minuttia, Optimist, Rock The Rankings and Skale. On four of those six a handful of pages still came back unreadable, so only Graphite and Rock The Rankings were read end to end. The other fourteen are larger than the 200-page-per-agency cap, so for those the corpus is the highest-priority 200: case studies, results and customer-story pages first, then pricing, then service pages, then everything else. We discovered 5,232 URLs on Directive alone and 3,255 on Single Grain, mostly blog. The "Pages read of discovered" column gives the readable count against every URL we found on that domain, so a row says what it rests on rather than implying a full sweep.

One agency read badly. First Page Sage returned only 101 readable pages from 200 fetched; the rest were blocks or redirects. Its row is the thinnest evidence here and is labelled with the number.

What we counted. A booked call, a demo request, a product signup, a qualified lead, qualified pipeline, or closed revenue, with a number on it. A percentage or a multiple counts. An email or newsletter signup does not, because it is not a buying step. Traffic, traffic value, citations, share of voice and visibility count appearances rather than commercial events.

The client has to be named. A figure with no client behind it cannot be checked by a reader. iPullRank publishes $2.4 billion in incremental revenue against a client it calls one of the world's biggest banks. Optimist publishes 49x growth in LLM referral revenue against "a large B2B technology company" in its AEO case study. Neither can be looked up, so neither passes.

And the client has to be someone else. LoudFace publishes a 2.0x lift in its own booked calls, counted from the Cal.com webhook and credited to its own AI answers. It is us measuring us, and an agency's own marketing is not a client outcome.

Only the agency's own pages count, and only about its own clients. Several of these agencies publish listicles ranking each other. Animalz writes about Skale, Minuttia writes about Breaking B2B's client, Skale writes about DerivateX's. A sentence about a rival's result is not the agency publishing its own. Each row names the page or pages that decide it, and a checker refuses to print a row unless every quoted sentence is in the fetched body of the page it sits beside.

What a verdict is not. It is not a check of whether the number is true. We did not audit anyone's analytics. We checked what each agency chose to publish and what it credited the result to.

What a buyer should ask

The table says who publishes what. It does not say whether the number survives a question. Four to ask on the first call:

1. What was the event? A booked call, a demo, a signup or closed revenue. A visibility percentage is not an answer.
2. Which channel got the credit? This is where most of the field sits. A lead figure credited to organic search is a real result and it is not evidence about AI search. Ask for the split.
3. How was it attributed? Tagged links, a self-reported source field, or a path traced in the CRM. Each method has holes. An agency that names its holes is telling you the truth.
4. Which client, and can I read the page? If the answer is an industry label, you cannot check it and neither could we.

The 14 questions to ask an AI search agency before you sign covers the rest of the contract. How to verify an AEO agency's results before you hire one walks through checking a claim yourself, and how to measure AEO agency ROI covers the attribution mechanics. If you want GEO agencies ranked rather than audited, the GEO agency shortlist ranks 24 of them for B2B SaaS. And what 192 AI answers actually retrieve for B2B SaaS agency questions classifies every page the engines cited by page type.

What this audit does not settle

Pages change, and this corpus is one day: 20 September 2026. The manifest carries a sha256 for every page, so a re-run tells you exactly which pages changed rather than leaving you to guess. Ask us for it and we will send it.

It does not settle whether crediting a result to AI search makes an agency better at the work. Six of the twenty publish commercial numbers that go to organic search, that do not separate AI from it, or that name no client, and some of those are the largest figures in the table. What it settles is narrower and still useful: which agencies have committed a named client's commercial outcome to AI search on a page you can open, and which have committed it somewhere else.

Our own answer to the question about which channel gets the credit is on the AI search service page: visibility, share of answer and position per prompt, reported weekly with the actual answers linked, and engagements from $5k per month.