How to Choose an AI SEO Agency

Twelve questions split four ways, three per area. Each answer has a correct shape: a named method, a date, a number, or a document. An adjective in place of one of those four is the signal.

Measurement method and cadence

1. What measurement method do you use, and what does that method count?
2. What is the reporting cadence, and what arrives inside each report?
3. What counts as a citation in your numbers — a linked source, an unlinked brand mention, or both?

Baseline capture before work starts

A baseline is a recorded measurement of current AI-answer visibility, taken before any change ships. A baseline enables proof.

1. What does my baseline measure, and on what date do you capture it?
2. What is my starting citation rate, and across which prompts?
3. What happens to that baseline when an assistant changes model mid-engagement?

Assistant coverage and method

1. What assistants do you track, and how many of them?
2. What prompt set do you run, how many prompts does it hold, and who writes it?
3. What do you do when one prompt returns two different answers?

End of the contract

1. What do I keep when the engagement ends?
2. What is the minimum term, and what is the notice period?
3. What happens to the tracking account and the prompt set after the final invoice?

A guaranteed citation. Citation rate is the share of tracked prompts whose answer names or links the site. No provider sets that share, because the assistant selects its sources at answer time from an index outside every agency's control. A guarantee indicates a provider not measuring, and the same applies to a guaranteed ranking.

A missing measurement method. An agency must publish a measurement method before it quotes a price: the prompt set, the assistants tracked, the capture frequency, and the definition of a hit. Four unnamed components leave four unaudited gaps inside any number the agency later reports.

A promise of training-data placement. A provider offering to put a business inside a model's training data describes an outcome no on-site work produces. The next section states the mechanism behind that limit.

Two distinct mechanisms produce a mention in an AI answer: parametric memory and retrieval.

Parametric memory is what a model retains from its training corpus, fixed in the model's weights at training time. Entry into that corpus comes from years of independent coverage across the open web — press, directories, forums, reference entries, and third-party writing that no retainer commissions. No on-site change writes a business into it, and no agency schedules a training run.

Retrieval is the second mechanism: the assistant issues a query at answer time, fetches live pages, and quotes from what it fetches. Retrieval responds to crawler access, page structure, extractable answers, and entity clarity — four things an agency does change. Those four changes register within weeks and stand up to weekly measurement.

Published small-business retainers run $1,500 to $5,000 per month. Mid-market retainers run $5,000 to $25,000 per month. Project engagements run $5,000 to $50,000. Hourly consulting runs $50 to $300 per hour. WebFX publishes all four ranges; RevvGrowth and Fuel Online publish overlapping 2026 figures on pages selling retainers.

Three limits apply to all four ranges: seller incentive, national weighting, and self-selection. The numbers come from agencies with reason to anchor high. The numbers describe national mid-market scope rather than a single-location business. Most agencies publish no price, leaving a visible set weighted toward providers comfortable showing a large figure. Treat the four ranges as a frame and ask each provider for a written scope against a written number.

A prompt set is the fixed list of questions run against each assistant on each capture date. Verification takes three named items and two actions.

• Request the prompt set as a file, with every prompt readable.
• Request the capture date of the claimed result and of the baseline before it.
• Name the assistant and the answer surface: ChatGPT, Perplexity, Google AI Overviews, or Gemini.
• Run four of those prompts yourself in a single session and read the sources each answer lists.
• Compare your own result against the screenshot supplied, prompt by prompt.

Five checks take under an hour. A claim surviving all five is a measurement; a claim failing the first is a screenshot. A provider declining the prompt set has answered question eight of the twelve. Buyers wanting the same five checks run against their own domain can book the examination. The twelve questions above apply to this firm as well, and who you are actually hiring is answered on the record rather than in a sales call.