How we verify

1. Who gets listed at all

We list AI-specialist agencies only. On a general directory the specialist and the everything-shop appear in the same list, which makes the specialist invisible — so the first thing this site does is turn most applicants away. Of the firms we have examined, we list 6 of 15.

There are three routes in. A firm needs one.

1. AI work is 70% or more of the business

   Measured against the firm’s own public materials, not the figure it gives us. A declared share is multiplied by a credibility factor and capped near what the record supports; where the two disagree, we publish the gap rather than picking one.

2. Or 2 of 5 AI-first signals, each with a public URL

   A pure percentage floor would empty this category: almost every good AI SEO firm is an SEO firm that went AI-first. So the floor stays strict and this route stays generous in reach — but a signal without a source does not count, which is what stops the route from becoming a checkbox.

   • A named, publicly documented AI/GEO methodology the firm authored — not a blog post about someone else's.
   • AI tooling the firm built and operates itself, publicly demonstrable.
   • The service menu itself is AI work only — no separately sold non-AI service lines.
   • A publicly named AI-visibility measurement stack, i.e. they can show what they claim to move.
   • Named individuals whose stated roles are AI-specific, not generalists with an AI title.

3. Or the firm is AI-native or AI-vertical

   A company whose own product is the AI has no meaningful “share of AI work”. An agency that does fairly ordinary marketing exclusively for AI companies is the specialist a buyer in that market is looking for. Both are AI-specific in the way that matters; neither survives a percentage test. Every competitor collapses these into “AI agency” and loses the distinction.

We publish what we turn away, with the measured figure and the sources — but only when we actually measured it. “We could not check” is never published as “they failed”. That distinction is the difference between a directory and a liability.

Check yourself against this rule →

2. What we check

Two ladders. A base ladder every firm faces — is this a real, trading, honest business — and a category ladder that differs for every category we cover. The AI SEOladder asks things an AI PPC firm would find irrelevant, and vice versa. A large generalist structurally cannot pass a specialist ladder, which is the point.

1. Is AI search actually their business?

   • AI search is a named service, not a blog topic — I look for a dedicated AI search or GEO service page, not a paragraph about AI bolted onto the traditional SEO page.
   • When the AI search page first appeared — I check the archive for the first time that page existed. Conviction before 2024 reads differently from a page that went up six weeks ago.
   • The traditional-SEO vs AI-search split is knowable — I work out from their own site how much of the practice is AI search and how much is classic SEO wearing a new label.

2. Can they measure the thing they sell?

   • A named AI-visibility measurement stack — I check whether they name how they measure AI visibility at all — a tracking product or their own tooling. Anyone selling AI search visibility with no stated measurement method is selling a feeling.
   • A defined AI-search KPI — I look for a stated metric — citation share, prompt coverage, answer presence. Rankings is not an AI-search KPI.
   • A case study reporting an AI-search number — I check that at least one result is an AI-search figure rather than an organic-traffic figure. Selling AI search and reporting Google traffic is this category's signature swap.

3. Who would actually do the work?

   • A named search lead with a traceable public record — I look for a real person with published work, talks or studies attached to their name — not a team of experts.
   • Team size against client count — I compare the team page against the client wall to see whether the ratio is plausible.

4. What does an engagement cost, and how do you leave?

   • Stated pricing model — I look for a named model — retainer, project, performance or hybrid — rather than one hidden behind a call.
   • Discoverable contract length and exit terms — I check whether the commitment and how to leave are findable without signing anything first.

Five outcomes, never flattened into a pass rate

Verified

We found it in public record.

Stated

The firm says so and we have not corroborated it.

Pending

We have not run this check yet. Our backlog, not their failing — so it stays in the denominator.

Unavailable

We tried and the source blocked us. This never renders as a failure.

Not applicable

No such source exists here. Leaves the denominator entirely.

Conflating “a bot wall blocked us” with “this agency failed” is how a verification engine becomes a defamation problem. The five labels exist so that cannot happen by accident.

3. The TF Score

It measures how complete and how well corroborated our evidence is — not how good the agency is.A firm can be excellent and score low because we have not established much about it yet. We say so on every page the number appears on, because a 0–100 number next to a company name will be read as a rating unless it is relentlessly explained.

60 points are universal and 40 are specific to AI SEO.

2 pillars are declared but not yet scored, because we have no checks behind them yet. They are removed from the denominator rather than scored zero — counting an unasked question as a failure would defame every firm equally, and dropping it silently would inflate every score equally.

What a claim is worth

Evidence ages. Each check carries its own shelf life, and a result past it is worth less than a fresh one — a verified fact from three years ago is a historical note.

Bands

• Well evidenced — 80 and above
• Evidence-backed — 65 and above
• Developing evidence — 45 and above
• Limited verified evidence — 0 and above

How the order is decided

Firms are ordered by evidence confidence, then by category delivery record, then by how recently we checked. Scoring and ranking are separate on purpose: one number cannot honestly mean both “our evidence is strong” and “look at this firm first”.

Ranking is never for sale. The scorer selects from a database view that physically cannot see payment, subscription, claim status, featured slots or reviews, and the deploy gate asserts the order is byte-identical across every billing state. Sponsored placement exists, is capped, sits above the list, is labelled on every card, and never enters the structured data an AI assistant reads. Paid features are presentation only.

4. Proof tiers

Separate from the score: how strongly a firm’s own claims are established. Tiers are derived in code from what reviewers record, never set by hand — a badge an administrator can toggle is a badge for sale.

Publicly verified

We checked this firm's core claims against sources anyone can open, and they held.

Confidentially reviewed

A reviewer saw evidence for this firm's core claims that cannot be published — client work under NDA, usually.

Self-reported

These are the firm's own statements, published as such. We have not been able to corroborate them yet.

Insufficient public evidence

There is not enough on the public record to say much either way. That is a statement about the evidence, not about the firm.

5. Tools

A firm may tell us what is in its stack and we label that as its statement. Calling a tool a specialism is a claim about capability, so it needs corroboration; without it we show the tool as a stack entry and say on the profile that we walked the claim back. Declaring tools cannot move a firm up any list.

6. Getting it wrong

We will. When we do, correcting a factual error is free and open to everyone — claimed profile or not, paying or not. Charging a company to fix something untrue we published about it is not defensible in any of the countries we operate in. Only the promotional right of reply sits behind a claimed profile.

Report something wrong →

Reviews

Reviews never affect a firm’s position and never gate a badge. They are a paid-tier feature, so either would make both purchasable by proxy. We do not reprint competitors’ review scores — those are collected for their platforms, not ours. Where a firm has a profile elsewhere we say so and link it, and leave the number where it was earned.