The short answer
Ask for the measurement frame before you ask for anything else. A real AI visibility programme runs a fixed set of buyer prompts across every answer engine, every month, and hands you the raw transcripts. An agency that cannot describe its prompt set, its engine list and its reporting cadence in one sentence is selling you ordinary SEO with a new cover.
Everything below is a way of testing that one claim.
The nine questions
Send these before the first call. The answers, or the absence of them, will tell you more than any deck.
- How many prompts are in the set, and do they ever change? Expect a number between 100 and 300, and a firm no on changing them. Editing the prompt set between reports is how a programme manufactures progress.
- Which engines do you measure? ChatGPT, Google AI Overviews, Google AI Mode, Perplexity, Gemini, Claude and Copilot is the current full set. An agency measuring only ChatGPT is measuring a fraction.
- Do I get the transcripts or a summary? Transcripts. If you cannot read the answer the model gave, you cannot check the number in the report.
- When I am not cited, do you tell me who was? The competitor holding the slot is the most useful line in the whole report, and the one most often left out.
- What is your plan for third-party corroboration? This is the heaviest lever for ChatGPT and the slowest to move. An agency with no answer here is working only on your website, which is the easy half.
- What will you not guarantee? Any agency that guarantees placement in a model output is describing something nobody controls. The honest answer names the limit.
- Who owns the work? Code in your repository, ad accounts and analytics in your name, documentation you could hand to someone else. Anything held in an agency account is leverage over you later.
- Is llms.txt part of your pitch? If it is the centrepiece, walk. No major provider has publicly committed to reading it in production and Google has said it does not support it. It costs an hour and it is not a strategy.
- Who checks the facts in the content? A named person, before publication. Models fabricate specifics with complete confidence, and a fabricated statistic in your name is a liability, not a ranking.
What a real programme includes
Whatever it is called, the work divides into four parts. An agency that covers only the first two is doing technical SEO.
| Part | What it means | How to verify it happened |
|---|---|---|
| Entity clarity | One canonical description of who you are, identical across your site, your structured data and your profiles | Ask to see the entity model as a document |
| Structured data | A JSON-LD graph cross-linked by identifier, not isolated fragments per page | Every reference should resolve to a defined node |
| Quotable content | Complete answers in the first two sentences, question-shaped headings, visible dates | Read a page they published and try to lift one sentence from it |
| Off-site corroboration | Mentions, reviews, comparison listings and community presence | Ask for the target list and the progress against it |
The fourth is the one that separates programmes. It is slow, it is unglamorous, and it is what the models weigh most heavily.
Red flags
- A guarantee of placement. Outputs vary by prompt, session, region and model version. No provider sells position in an organic answer.
- A prompt set that changes each month. The number will always go up. It will mean nothing.
- Reporting without transcripts. An unverifiable number is a feeling with a percentage sign.
- llms.txt as the headline. The cheapest item on the list, sold as the whole list.
- No named person on the content. Google AI Overviews weighs expertise and authorship. Anonymous content is a weaker signal, and nobody is accountable for the claims.
- Work held in agency accounts. If leaving costs you your history, the retainer is being protected by lock-in rather than by results.
- One engine reported as "AI search". Only around 11% of domains cited by ChatGPT are also cited by Perplexity. One engine is not the channel.
What it should cost
Very few agencies publish, so there is little to compare against. As a reference point, these are our own published numbers, and they are in the range we would expect from anyone doing the work properly.
| Engagement | Range | What drives the number |
|---|---|---|
| AI visibility retainer | $1,950 to $4,200 a month | Size of the prompt set, number of engines, how much content is produced |
| Full multi-channel retainer | From $8,500 a month | Paid, social and lifecycle running alongside, plus one measurement layer |
| AI-optimized website | $4,900 to $30,000 one time | Page count, integrations, whether it is a marketing site or an application |
| Paid media management | 10% to 15% of spend | Usually with a monthly minimum around $1,800 |
Treat a quote well under these as a signal to ask what is missing, and a quote well over them as a signal to ask what is included. Ask for a minimum term as well as a monthly figure. Six months is reasonable for AI visibility, because corroboration is slow. Twelve is not, and anything with no exit is a warning.
Can your current agency do it?
Sometimes, and it is worth asking before you go looking. The technical foundation of AI visibility is shared with technical SEO, so a strong SEO team already owns half the work. The half they usually do not own is the measurement frame and the off-site corroboration plan.
Send them the nine questions. If they come back with a prompt set, an engine list and a corroboration plan, keep them. If they come back with a proposal to add llms.txt and some FAQ schema, they are guessing.
How to decide
Ask three agencies for the same thing before you ask any of them for a proposal: run twenty of your buyer prompts across the engines they claim to measure, and send you the raw transcripts.
It costs them an afternoon. What comes back tells you whether they have the tooling, whether they understand your category, and whether they will show you evidence or a summary. The one that sends transcripts with the competitors named is the one to brief.