I ask three models - Claude Sonnet 5, Gemini 3.6 Flash and GPT-5.6 Terra - the same discovery questions I ask for every tool in a category, with web search turned off, and I record which tools they name back. A prompt looks like a real buyer question, for example: "I want to build a video meetings app. What's a good no-code solution for it?" There are twelve of them per category, covering different buyer intents, and each one is asked five times of each model: 180 answers per category. Asking once tells you very little, because these models don't answer identically twice.

The score combines how often a tool is named with where it lands. Named first counts for full credit and each position below that counts for less, on a fixed curve that is the same for everyone. Same prompts, same models, same maths, every vendor in the category, one sitting. It's not weighted by who I know or who I like.

The exact prompt wording stays private. Publishing the bank would just hand everyone a list to optimise against, and then the score would measure who read this page rather than what the models know. The method is public; the questions are not. If you think your own number is wrong, ask me and I'll send you the row-level data behind it - every model, every question, whether you were named and in what position.

I decided to disable web search because SEO we all do and work hard to optimize. Having the original LLM incorporate the knowledge of a tool into its training is a lot more powerful than when it deploys search to gain the data it needs.

Three different things get called "found via AI", and only one of them is what this column measures:

1. What the model already knows - it answers from training, no browsing and no searching. That's what AI Visibility measures here
2. What the model retrieves when web search is on - now it's reading live pages in the moment. Different mechanism
3. Referral traffic your analytics label as "AI" - someone clicked through from an AI product

You can be strong on 2 or 3 and still score low on 1. A low number here means the models don't reliably recall you unprompted - not that nobody ever reaches you through an AI tool.

ahrefs measures the web link graph - who links to whom, domain authority, classic SEO. AI Visibility measures what the model absorbed in training and recalls without retrieval. Those two can diverge hard. You can have solid domain authority and still be near-invisible in a model's memory, because recall tracks how often and how consistently you're mentioned across the training corpus, not your backlink profile. The ahrefs chart is measuring a real thing. It just isn't the thing this column measures.

The general direction is no secret: get mentioned, consistently and in context, across the kinds of sources these models train on - and make yourself easy for an agent to read. It's a real, teachable discipline, not a backlink campaign. It's also the slowest and hardest column to move: changes you make now only show up when the next training cycles absorb them, so you're looking at least at a month or two before you know if something worked. I'm working through the same thing on my own products, so I'm not speaking from the sidelines.

Turning that direction into a prioritized, specific plan for your product - where your gaps are and what to do first - is where I come in as advisory. The method is public; the custom plan for your own product is the work I do. Start that conversation here.

Quality is a published rubric - documentation and feature coverage against a fixed checklist. The checklist is specific to each category and lists the must-haves and nice-to-haves for that category (the list itself was built by having AI survey the vendors' sites in the category and surface what the field offers). I didn't hand-grade anyone. I asked Claude to read the docs and score them against that checklist - the same thing, the same way, for every vendor in the category in one sitting.

The checklists and what each vendor scored against that checklist is proprietary. Two reasons why: a public per-vendor scorecard turns every checkmark into a negotiation, and the detailed "here's exactly where you lost points and what to fix first" is the advisory work, not the public layer. If you want your own breakdown, that's a conversation - start it here.

Two layers, kept separate on purpose. The measured columns - AI Visibility and Quality - are mechanical: same prompts, same rubric, same models, applied identically to everyone. The editorial layer - the badges and the short take - is my read, and I own that as opinion. The numbers aren't a matter of taste.

No, and no. Inclusion isn't paid, and no score can be bought. If you work with me on improving your standing, that advisory helps you understand and close your gaps - it never moves your number on this hub. The scores stay independent of who pays me for anything. That independence is the entire reason a score here is worth something - to you as much as to a buyer reading it.

Yes, always. Facts aren't a judgment call. Tell me the specific - a wrong link, a feature I misread, pricing that changed, or a name that should read differently - and I'll correct it. The score is my read; the facts behind it should be right.

Maybe, over time. Some of these are genuinely useful buyer criteria. But I'm not going to bolt on a new scoring dimension the same week a vendor who'd benefit from it asks for one - you'd want me to hold that line for you too. Fair asks get logged and considered for a future revision.

Claude Sonnet 5, Gemini 3.6 Flash, and GPT-5.6 Terra, as of the August 2026 run. The exact versions are named because model recall shifts between versions - when I re-run discovery on newer models, the version strings roll and the scores move with them. I pick the current mainstream default of each provider rather than the biggest or the cheapest, because that is what the models are actually answering your buyers on.

The panel before this one was Claude Sonnet 4.6, Gemini 2.5 Flash and GPT-4o. Old and new models are never mixed inside one score: when the panel rolls, every published category is re-run on the new panel in the same cycle.

It measures how ready a tool's WebRTC surface is for an AI agent to actually use it - operability, not marketing. The tiers run from nothing, through a generic then a hand-authored llms.txt, up to a real mechanism the agent can drive: an MCP server that calls the API, or executable agent/operator skills. A docs-only surface, including an MCP that just serves documentation, sits below one that can operate the product, and tooling that does not reach the WebRTC surface does not count. It's an early, deliberately strict measure - most of the field scores low today, which is the point.

Like other measurements and scoring here - I didn't manually check - I asked Claude to figure out if he can easily find these or not.

There are two clocks, and which one you're on depends on what changed.

Facts move monthly. Prices, product names, ownership, whether you're on the list at all. If something on your row is wrong, tell me and it gets fixed at the next monthly pass, sooner if it's materially wrong and live. Nothing is re-scored, so there's nothing to wait for.

Measured things move quarterly. AI Visibility, Quality, agent-ready tier. Every published category is re-measured together, in one cycle, on one model panel: February, May, August and November. The dateline on each category page tells you which snapshot you're looking at, dimension by dimension.

The reason measured scores don't move on request: a score only means something against the other scores measured the same day, with the same models, from the same questions. Re-running one vendor because they asked would hand that vendor a fresher measurement than everyone they're ranked against, which is exactly the sort of thumb on the scale this guide exists not to have. So the answer to "can you re-check us now?" is no, for the same reason it will be no when your competitor asks.

Worth knowing which lever you actually pulled: docs, an llms.txt and an MCP are things an agent can find today, so those show up at the next quarterly pass. AI Visibility is the slow one - it moves when the models' training catches up with the world, not when you ship.

Because the model's memory shapes the search. When someone asks an AI for a shortlist, the model's prior knowledge decides which names it reaches for, which sources it trusts, and how it frames the comparison - before and during any retrieval. And plenty of AI usage still runs with no browsing at all: API calls, coding agents, embedded assistants. A tool the models already know gets pulled into answers everywhere; a tool they don't is betting everything on winning the live search every single time. That's why I measure the memory layer on its own.

No. The notes column is my read, not vendor copy. If your price is wrong on the page, that's a fact and I'll fix it the moment you tell me. But a comparative claim - cheapest, fastest, better, best, bestest, awesomeful - is marketing, and the whole value of this hub is that nothing in it was written by the vendors being scored. The same wall that keeps your competitors' claims out keeps yours out.

Placement follows the product, not the press release. A product that's still sold, documented and running stays listed as long as buyers can still choose it - with my read on its trajectory where that's warranted. Category assignment follows product heritage: where the product actually competes, not where the parent company would like to be filed. When a product is finally gone, it comes off at the next refresh.