Does AI name you when nobody says your name?
Ask an engine about your brand and it will talk about your brand — that measures nothing. AEOMAX asks the question your buyer actually asks, with your name left out of it, and counts how often you come back, five draws at a time, so one lucky answer never reads as a trend.
Then it shows you the sentence, the page the engine took it from, and what the fix was worth — in sessions, conversions and revenue. Every figure on every screen carries its sample size, because a number that will not say how sure it is is an opinion in a suit.
The engines coverage
Six engines, on every plan, and we say how each one is read.
Included, not metered one paywall at a time. The same question goes to all six in eight markets, written in that market's own language, because the brands named in the US are not the brands named in Germany or Japan. Every reading records whether the engine searched the web first or answered from memory.
Google AI Overviews has no API, so it is read from the rendered results page — a run that is blocked or served a CAPTCHA is excluded, never scored as a zero. We do not sell Microsoft Copilot or Meta AI: neither ships a public consumer-assistant API, so nobody can honestly measure them. Read the methodology.
The lead measure unaided recall
Take your name out of the question.
Almost every AI-visibility tool asks the engine about your brand. Of course the answer mentions your brand — you put it in the question. That number goes up and to the right and means nothing.
The unaided index asks what your buyer asks: best CRM for a small sales team. No brand name. It asks five times per engine per cycle with the answer cache bypassed, because one draw from a model is an anecdote. Then it reports how often you were named as x of k, with a 95% Wilson interval — so next cycle you can tell a real move from noise.
Off until you have written ten category questions, and five draws per question is fixed. Available from Growth.
Unaided recall · 30 days
best CRM for a small sales team
CRM that works with HubSpot
cheapest CRM for under 10 seats
CRM with the best mobile app
CRM for a team leaving spreadsheets
Sample shape, not a real brand: 5 questions, 9 answers each, 45 in all. The last row is the one worth reading — asked nine times, named by nobody.
The method four stages
From invisible to the default answer.
Every screen belongs to one of four stages. You always know which question you are answering, and what to do next.
Stage 1
Diagnose
Why am I invisible?
- ✓Engine breakdown
- ✓Site health
- ✓AI access firewall check
- ✓AI crawl readability
Stage 2
Find openings
What do I target next?
- ✓Prompt winnability
- ✓Prompt discovery
- ✓Citation gap
- ✓Knowledge vs retrieval
Stage 3
Compete
How do I pass rivals?
- ✓Share of voice
- ✓Battlecards
- ✓Rival move radar
- ✓Shopping visibility
Stage 4
Prove it
Did it pay?
- ✓Per-prompt revenue
- ✓Impact ledger
- ✓Lift experiments
- ✓Client-ready PDF
Why this journal six editorial positions
Most tools count mentions. We tell you what they are worth.
Accuracy
An LLM reads every answer
Keyword matching cannot tell a recommendation from a brush-off, and calls most answers neutral. A model reads each response: are you genuinely recommended, where do you rank, how are you described — and does the answer state anything about you that is simply wrong?
Hallucinations
Catch what AI gets wrong, and watch it persist
Dead pricing, a feature you retired, a plan that no longer exists. Register what is true, and every contradiction becomes one row in a ledger: which engines repeat it, how many runs, and how many days it has survived since you first saw it.
Revenue
Visibility tied to money
Connect Google Analytics and see AI-referred sessions, conversions and revenue split by engine and by day — measured. The split down to individual prompts is modelled from which answers cited the page people landed on, and is labelled as modelled wherever it appears.
Honesty
Thin samples say so instead of rounding
A score from 8 answers and one from 800 look identical everywhere else. Every figure here carries the sample it came from, mention and citation rates carry a 95% Wilson interval, and anything under five readings is labelled insufficient rather than dressed up as a number.
Provenance
Which sentence a cited page actually repeats
Take one archived answer and see, sentence by sentence, which cited page repeats its wording — with the passage quoted. Sentences nothing matches are listed separately from pages we could not read. It is word-overlap, not meaning: a paraphrase of a real source shows as unmatched, so treat it as a floor.
Action
Fixes you can ship, then measure
The plan ships the artefact to copy out — llms.txt, Organization and FAQ JSON-LD, robots rules, content briefs — and llms.txt can go to your GitHub repo as a pull request you review and merge yourself. Mark a fix shipped and the ledger reports the 14- and 28-day change either side of that day, flagged if an engine model update landed inside the window.
Everything it prints the full run
Six things, end to end.
Measure the answer, take it apart, fix it, earn the citations behind it, prove what it was worth — and run the whole thing for a book of clients.
Measure
What every engine says, every day
The same buyer questions, asked of every engine on a schedule, read by a model rather than a keyword match.
- Visibility score, share of voice, position and sentiment per engine
- Grounded-or-from-memory label on every engine, from the stored answer
- Blocked pages and CAPTCHAs excluded instead of scored as a zero
- Ask a prompt up to 5 times where you turn it on: “mentioned 2 of 3”, with the spread
- Engine model changes noted on the trend by us, so a drop is not misread
- Sample size on every figure, and a 95% Wilson interval on the rates
Read the answer
Not just whether — what, and from where
The full text of every answer is kept and hashed, then taken apart: what changed, what it came from, what is simply wrong.
- Answer change feed: the exact words each engine changed, run over run
- Provenance: which sentence a cited page repeats word-for-word, and which nothing does
- Knowledge vs retrieval: does the model know you, or only find you
- False-claim ledger: one row per thing AI keeps getting wrong, and for how long
- Battlecards: the objections engines raise about you, with the answer behind each
- Paid vs earned share where the answer surface carries ads
Act
A fix with an owner, not a chart
Every finding becomes a card someone owns, with the artefact attached and the result measured afterwards.
- Work Loop board: owner, due date and a history that survives every rescan
- Ship-ready llms.txt, JSON-LD, robots rules and content briefs to copy out
- Answer Deploy: a pull request in your GitHub repo with that llms.txt — we never merge it
- Linear sync: move a card to In progress and the issue is created, closed status comes back
- Impact ledger: the 14- and 28-day change either side of the day you shipped
- Winnability score and lift experiments with a holdout arm
Earn citations
The sources the answer is built on
AI answers quote a handful of sites. These reports name them, type them, and track your way onto them.
- Cited sources, typed as review / forum / docs / news / owned
- Source value index: which kinds of source travel with a mention
- Citation gap: sites cited in your category that never name you
- Publisher graph, down to the named byline where a key for it is configured
- Reddit threads and review platforms the answers lean on
- Outreach pipeline with the citation delta after you land
Prove it
Visibility, in money
Connect analytics and the score stops being a vanity metric — down to which prompt the revenue came from.
- AI-referred sessions, conversions and revenue per engine, by day
- Per-prompt revenue — modelled from which answers cited the landing page, not observed
- Search Console joined in: impressions holding while clicks fall, stated as correlation
- Agent traffic split by bot: training crawl, search index, live agent
- Crawl-to-refer: which engine crawls you daily and sends nobody back
- HubSpot: AI-sourced pipeline back, visibility score written on the company
Run the business
Agencies, teams and procurement
Every client brand in one place, your branding on the way out, and the answers an enterprise buyer asks for.
- Portfolio across every client brand with its 7-day move and open alerts
- Pitch a prospect on credits without spending a paid brand slot
- White-label dashboards, revocable client links, scheduled branded PDFs
- Partner program: provision client workspaces on negotiated limits, one invoice
- Per-brand access, seats, invitations and an append-only audit log
- SSO and SCIM, SIEM streaming, and a published trust and subprocessor pack
Anything that depends on a third party — Google Analytics, Search Console, HubSpot, Linear, GitHub, Trustpilot, Reddit — says plainly that it is not connected until you connect it. Where a screen shows a sample shape instead of your data, it says so on the screen, every time.
On the record trust & method
Built to be checked, not just believed.
We publish how the numbers are produced — which surface each engine is asked on, whether it retrieved or answered from memory, how many responses sit behind a figure and how wide the interval is. If the sample is too thin to support a claim, the product says so instead of printing a number.
Stated methodology
Wilson intervals for rates, standard error for scores, sample size on every metric.
Access your auditor can read
Roles and per-brand grants, an append-only audit log, SSO and SCIM, and SIEM streaming into your own Splunk or Datadog.
Procurement pack, checked against the code
Subprocessor list, data-flow diagram and retention schedule — with a test that fails the build when the published retention no longer matches what the database actually does. Read it at /trust.
The agency desk clients at scale
Ship it under your name
Client-ready PDF
Your organisation in the masthead, selectable text, working links.
Portfolio & pitch
Every client on one screen with its 7-day move — and scan a prospect on credits before they sign.
Alerts where you work
Slack, email or any webhook the moment visibility slips.
Scheduled & partner-ready
Weekly PDFs in each client’s inbox, and a partner program that provisions client workspaces on negotiated limits.
Put AI visibility into your own stack.
Everything on the screen is also an endpoint, an event or a file in your warehouse — and an MCP server your own AI assistant can read.
- ✓REST API with scoped aeo_ keys and a versioned OpenAPI contract
- ✓MCP server, read-only — ask Claude or Cursor about your own visibility
- ✓Signed webhooks with retries; Zapier, Make and n8n triggers
- ✓Nightly export to your own S3-compatible bucket, to load into your warehouse
- ✓Slack app with slash commands, and a Looker Studio connector
- ✓Your own verified hostname serving llms.txt and a read-only MCP endpoint — experimental
Example
GET /api/v1/public/brands/:id/scores
{
"data": {
"brandName": "Acme Corp",
"platforms": [
{ "platform": "chatgpt", "avgVisibilityScore": 72 },
{ "platform": "perplexity", "avgVisibilityScore": 84 }
]
}
}Free editions read before you pay
See the number before you buy anything.
Live demo
Read a whole workspace, filled in
Every screen you would get, populated for a fictional brand: the score with its interval, all six engines, the prompt table, the work loop and the revenue reconciliation. Sample data, clearly labelled, no account.
Open the demo →Free grade
Grade your domain, no account
Type a domain. Real engines answer real buyer questions about it, and you get the score, the gaps and the first fixes. Three free grades per visitor a day.
Run a free grade →Public index
The AI Visibility Index
Which brands AI assistants name in a category, ranked, with the grounded-or-from-memory mode stated per engine and a dated methodology. Brands appear only when their owner opts in.
Browse the index →Your site
A badge you can embed
Verify your domain with a DNS record or a meta tag, then paste one line to show your live AI Visibility Score on your own site. Switch it off and it stops serving within five minutes.
Included from Starter.
The claim
Every other tool hands you a score. This one hands you the score, the sample it came from, and how sure it is — and if the sample is thin, it says so instead of rounding.
Six engines, asked the way a buyer asks — with your name left out of the question. A model reads every answer rather than counting keywords. Rates ship with a 95% interval, and a claim ships with the sentence and the page it came from.
How this is measuredAgainst the alternatives a fair fight
“My SEO tool added an AI tab. Why pay for this?”
A fair question, and for some teams the bundled tab is the right answer. Here is where the difference actually is, stated as what we do rather than as what anyone else does not.
| The question | The usual answer | Ours |
|---|---|---|
| Does AI mention me? | Ask the engine about your brand and count mentions. | Ask the category question with your name removed, five times per engine, and report x of k with a 95% interval. |
| How sure is that number? | One number, no sample size. | Every figure carries its sample; rates carry an interval; under five readings is labelled insufficient, not rounded. |
| Why did it say that? | The answer text, and a list of cited links. | Sentence by sentence, which cited page repeats the wording, with the passage quoted — and which sentences nothing matches. |
| Did the engine know me, or find me? | Not distinguished. | Every reading records whether the engine searched the web first or answered from memory, derived from what the answer did. |
| What is AI getting wrong about me? | Sentiment, positive or negative. | Register what is true; every contradiction becomes one ledger row with the engines that said it and how many days it has stood. |
| What do I do on Monday? | A list of recommendations. | A ranked plan with the artefact attached — and llms.txt goes to your repo as a pull request you merge yourself. |
| Did the fix work? | Watch the chart. | A treatment arm and a holdout you choose, read out with an interval — or an explicit “insufficient” with the sample sizes. |
| Was it worth money? | Sessions from AI referrers. | GA4 sessions, conversions and revenue per engine per day — measured — plus a modelled split to prompt level, labelled as modelled. |
If your team needs AI visibility as one line in a weekly SEO report, the bundled tab is cheaper and you should use it. This is built for the case where the number is going in front of someone who will ask how you know.
Subscription rates monthly, cancel any time
All six engines on every paid plan.
No per-engine add-ons. You pay for brands and prompts; coverage is never the upsell. The usage meter shows exactly how many AI answers you have used against your allowance, before any invoice.
Free
$0/mo
One brand, to see a real reading on your own domain.
- ✓1 brand
- ✓10 prompts
- ✓2 engines
- ✓Weekly tracking
- ✓7 days of history
Starter
$49/mo
For solo operators and founders.
- ✓1 brand
- ✓30 prompts
- ✓5 competitors
- ✓All 6 engines
- ✓180 days of history
Growth
$149/mo
For growing teams.
- ✓5 brands
- ✓150 prompts
- ✓15 competitors
- ✓Daily tracking
- ✓Unaided index · revenue · API
Pro
$349/mo
For power users and agencies.
- ✓20 brands
- ✓500 prompts
- ✓50 competitors
- ✓Geo + language markets
- ✓Lift experiments
Agency
$799/mo
For multi-client scale.
- ✓Unlimited brands
- ✓2,000 prompts
- ✓White-label reports
- ✓Webhooks + Slack
- ✓Role-based access
An answer is one engine reading one prompt once. Your allowance is brands × prompts × engines × cycles, and the meter shows it before any invoice — nothing past the allowance is charged unless you switch overage billing on yourself.
Letters the awkward questions
Letters to the editor.
The questions buyers actually arrive with, answered without a euphemism. The signatures are ours, not testimonials — nobody on this page is pretending to be a customer.
“How many prompts do I actually need?”
— Getting started, week one
The editors reply Ten category questions is the floor, because that is where the unaided index switches on. Thirty covers a normal buying journey — comparison, pricing, use case, integration, objection. More than a hundred is usually a sign the questions have stopped being ones a buyer would type.
“What happens when I hit my allowance?”
— Burned by metered billing before
The editors reply Nothing is charged. The meter shows your usage against the allowance all month, and past it scans keep running on a soft cap. You are only billed for overage if you switch it on yourself.
“Is my data used to train anything?”
— Legal, on the second call
The editors reply No. Your prompts go to the engines as queries and the answers come back to your workspace. The trust page names every subprocessor, what each one receives and how long each store keeps it, and a build-time test fails if that page drifts from the code.
“The engines change every month. Does that break the numbers?”
— Skeptical of all dashboards
The editors reply It moves them, which is why a model change is noted on your trend rather than left for you to discover. The change feed diffs each answer against the previous run for the same prompt and engine, so you can see whether the words changed or only the model did.
“Are you SOC 2 certified?”
— Procurement, question 14 of 40
The editors reply Not yet, and the trust page says so with a dated roadmap rather than a badge. It also publishes the gaps — what is encrypted and under which key, what has no rotation, what has never had an independent pen test. Read it before you decide.
“Can I cancel?”
— Asking for a friend
The editors reply Monthly, from the billing page, no call. Your data stays readable until the period ends and exports are available the whole time.
“Why should I trust your score?”
— A statistician, arms folded
The editors reply Do not. Check it. The methodology page states, per engine, which surface the reading came from and whether it searched the web. Every rate carries its sample and interval. Thin samples say insufficient. And the free grade runs on real engines before you give us anything.
“What does it not do?”
— Comparing three vendors this week
The editors reply It does not measure Microsoft Copilot or Meta AI — no public consumer API exists, so nobody can. It does not merge your pull requests. It does not randomise experiment arms, so a lift readout is a controlled before-and-after against a holdout you picked, not a randomised trial. Several capabilities need a connection — GA4, HubSpot, Linear, GitHub — and say so where they appear.
Managed service
Rather have it done for you?
Everything in Agency, plus the people. We run the playbook the reports point to and prove the result with the same visibility numbers you see in the dashboard.
- ✓A named AEO specialist who owns your AI visibility
- ✓Fixes shipped for you: schema, llms.txt, crawler access, content gaps
- ✓Citation outreach to the sources AI engines already trust
- ✓Monthly report: what shipped, and what it did to your visibility
Custom / scoped monthly
10–40 specialist hours a month, invoiced separately from your plan.
Request managed serviceFind out what AI says about you today.
Add one brand, tell us what you sell, and the first answers start landing while you watch. You get a ranked list of what to fix, with the artefact attached and the follow-up measured.
Not ready? Grade your domain on real engines, no account →