"AI SEO agency" describes two different services, and the label does not tell you which one you are being sold.
Model A uses AI internally to produce SEO work faster — research, drafting, clustering, technical analysis. The output is traditional SEO; the production method changed.
Model B optimises your business to be retrieved and cited by AI systems — AI Overviews, AI Mode, ChatGPT, Perplexity. The output is a different discipline; the destination changed.
Most brands being pitched "AI SEO" are being quoted for Model A at Model B prices. This article separates them, prices them, and gives you the questions that force a straight answer.
The two-service problem
The confusion is not accidental. Both services can honestly use the same three words, and the overlap in tactics is real enough that vendors can defend the ambiguity.
Model A — AI-assisted SEO production
Everything a competent SEO team already does, executed faster. Keyword clustering at scale, draft generation, internal-link mapping, log-file and crawl analysis, schema generation, reporting automation.
The value is throughput. A team that could brief eight articles a month can brief twenty. A technical audit that took two weeks takes three days.
The important thing to understand: this changes your agency's cost base, not your results mechanism. You are still competing for rankings, still measured on traffic and conversions, still subject to the same ranking factors. If the efficiency gain is real, some of it should reach you as either lower cost or higher output — and if a proposal offers neither, you are funding the agency's margin improvement.
Model B — optimisation for AI surfaces
Work aimed at whether AI systems retrieve, cite and correctly describe you. Retrievable content, corroboration on sources those systems draw from, consistent facts wherever your business is described, and a measurement layer built on prompts rather than keywords.
The value is presence in a surface your current reporting may not isolate. Google states that sites appearing in AI features are "included in the overall search traffic in Search Console," reported within the "Web" search type — so the traffic is counted, but it is not broken out.
Why the label hides the difference
Because both are true statements about a business using AI in SEO, and because Model A is far cheaper to deliver. An agency that has automated its production and an agency that has built a prompt-measurement practice can both say "AI SEO agency" without lying. Only one of them is doing work your incumbent is not.
What almost no standard retainer includes
Here is the part that matters commercially, and it is the reason this distinction is worth your time.
Most SEO retainers — including good ones — contain none of the three capabilities that Model B actually requires. Not because the agencies are weak, but because these capabilities did not exist as a discipline eighteen months ago and retainers do not spontaneously expand.
One — prompt-based measurement. Keyword rank tracking cannot tell you how a system describes your business when someone asks a question in your category. That needs a versioned prompt set, a fixed cadence, and a person reading the answers. Tooling exists — Semrush's AI Visibility Toolkit at $99/month with 25 tracked prompts, Ahrefs' Brand Radar from $50/month or $199/month against a 455M-prompt index — but the subscription is not the capability. Knowing which prompts matter in your category, and what to do when the answer is wrong, is the capability.
Two — the accuracy audit. What are these systems saying about you right now, and is it correct? This appears in no standard scope of work. It is the fastest source of fixable commercial damage we find, and in most accounts nobody has ever looked. Wrong service descriptions, superseded pricing, positioning a company abandoned two years ago — all being repeated confidently to buyers in the research phase.
Three — source-level corroboration targeting. Not link building for domain authority. Identifying the specific third-party sources cited in your category's AI answers and getting into them. Semrush surfaces these as Source Opportunities — sources cited in your category that mention competitors but not you. Ahrefs reaches the same target list through citation tracking. Neither is a link-building brief, and neither is in your current plan.
These three are the gap. A proposal containing all three is Model B. A proposal containing none is Model A with new vocabulary, and you should price it as production.
What the gap costs while it stays open
This is not a theoretical exposure. Three things happen in the interval.
Wrong answers compound. Every week an incorrect description of your business goes uncorrected, it gets corroborated further — repeated by aggregators, quoted in roundups, absorbed into the consensus these systems draw on. Corrections get more expensive the longer they wait, because you end up correcting thirteen sources instead of one.
Competitors take the category prompts. The prompts where a buyer has no brand in mind are the ones that introduce a business to someone who had never heard of it. Those positions are contested now, and the brands corroborated across the sources these systems favour are taking them while the question is still open.
Your own best content goes uncited. Plenty of businesses have the right answer on their site in a form nothing can extract — buried in a case study, phrased as marketing copy, locked in an image. The asset exists and earns nothing.
None of that shows up in a rankings report, which is why it stays invisible in a standard retainer.
How to run the diagnosis yourself
Before you brief a single agency, spend an afternoon establishing which problem you actually have. This costs nothing and it changes the conversation entirely, because you arrive with findings rather than questions.
Step one — twenty prompts. Write the questions a buyer in your category asks before they know any brand names, plus a few comparisons naming competitors, plus your own brand. Run them in fresh sessions with no history. Record what comes back verbatim.
Step two — three columns. For each prompt: did you appear, was what was said accurate, and which source was cited. Those three columns produce three different diagnoses.
Absent from category prompts means a visibility problem. Competitors are being introduced to buyers who have never heard of you, and no amount of production volume fixes it.
Present but described incorrectly means an accuracy problem — the cheapest and most urgent thing on this list. Wrong service descriptions and superseded positioning are being repeated to buyers in the research phase, and every week they go uncorrected they get corroborated further.
Present and accurate, but every citation points at someone else's page means a source problem. Your content is not the thing being used, even when the answer is about you.
Step three — check whether anything moves. Run the same twenty prompts a week later. These systems are non-deterministic, so single-run results mislead. Two runs tell you what is stable.
You now know which of the three capability gaps you need, and you can ask an agency to price that specific gap rather than a package. The full method is in our AI visibility tracking guide.
What each model should cost
No published rate card in this category is worth anything, because scope varies more than price does. What can be stated is the logic each model should price on, and the mismatch to watch for.
Model A should price on throughput. Cost per deliverable, or a retainer that buys a defined volume of output. If an agency has automated production, its cost per unit has fallen, and a proposal that reflects none of that is asking you to fund a margin improvement. Ask directly what their cost per deliverable was eighteen months ago and what it is now.
Model B should price on scope of surface and depth of off-site work. The number of engines measured, the size of the prompt set, and how much corroboration work is in scope. Measurement tooling is a real and checkable line item — Semrush's AI Visibility Toolkit at $99 per month with 25 tracked prompts, Ahrefs' Brand Radar from $50 per month for custom prompts or $199 for its 455M-prompt index. Ask whether that is passed through at cost or marked up.
The mismatch to watch for is Model A scope priced on Model B logic: a production retainer sold at a premium because the proposal mentions AI Overviews. The test is simple. Ask which deliverables would change if you told them you did not care about AI search at all. If the answer is "none," you are buying production.
Side by side
| Model A — AI-assisted production | Model B — AI surface optimisation | |
|---|---|---|
| What changes | How work gets made | What the work targets |
| Measured by | Rankings, traffic, conversions | Mentions, citations, accuracy, share of voice |
| Tracking unit | Keywords | Prompts |
| Off-site work | Links for authority | Sources cited in AI answers |
| Deliverable | More output, faster | Presence and accuracy in generated answers |
| Your incumbent probably | Already does this | Does some of it accidentally |
| Fair pricing | At or below current retainer | Additive to it |
Which one your business actually needs
Model A is a production question. If your constraint is throughput, that is a capacity conversation with whoever produces your content, and the efficiency gain should show up in your cost per deliverable.
Model B is a market-position question, and it is the one with a clock on it. Run the check in our AI visibility tracking guide — twenty prompts, one afternoon. If competitors appear in category answers where you do not, or your business is described wrongly, you have a Model B problem that rankings reporting will never surface and a production retainer will never fix.
Foundations still matter underneath both. Google's eligibility condition is that a page be indexed and eligible to appear with a snippet, and content locked in images is content nothing can use. A capable Model B engagement fixes that on the way through rather than treating it as a separate project — it is the same work, sequenced first.
Where the data disagrees
There is a genuine, unresolved dispute underneath Model A that buyers should understand, because vendors on both sides state their position as settled fact: does AI-assisted content production help or harm search performance?
Position A — the method is irrelevant, only quality counts. Google's guidance on AI-generated content has been consistent: it rewards high-quality content regardless of how it is produced, and the policies target scaled content abuse rather than automation as such. On this reading, an agency using AI to produce genuinely useful content faster is simply more efficient, and the output should perform identically.
Position B — scaled production degrades quality in practice. The counter-argument is not that Google penalises the method, but that volume-oriented production tends to produce commodity content, and commodity content underperforms. Google's own AI-features guidance emphasises "creating helpful, reliable, people-first content" and, in its optimisation guidance, warns against creating pages for every query variation primarily to capture generative surfaces. On this reading, Model A's throughput advantage is precisely the risk.
CONTESTED, and honestly so. Both positions cite the same documentation and read it differently.
Our read. The method is not the variable — the editorial standard applied to it is. AI-assisted production with a genuine research, verification and editing layer performs. AI-assisted production used to multiply output without that layer produces exactly the commodity content Google's guidance describes. The question to ask a Model A vendor is therefore not "do you use AI?" but "what is your verification step, and who performs it?" A vendor who cannot describe that step has automated the writing and not the thinking.
What would change our position: published data isolating production method from editorial standard at scale. Nobody has it, because the two are almost impossible to separate in the field.
What good looks like in a proposal
Six things separate a scoped engagement from a package. Most proposals contain two.
A baseline, taken before work starts. Current state across the engines in scope, using a documented prompt set. Without it, nothing afterwards is falsifiable and nobody can be held to anything.
A named prompt set that you own. Written down, versioned, and yours on exit. If the prompt set belongs to the agency, your measurement history ends when the relationship does.
Separation of on-site and off-site work, with an honest note on which on-site items are ordinary SEO. A proposal that cannot make that separation is either unsophisticated or hoping the question does not come up.
Named source targets. Not "digital PR" but the specific categories of source cited in your category's answers, and the order they will be pursued in.
A reporting model that shows losses. Ask to see a sample report. If every chart goes up and to the right, you are looking at a sales asset rather than a reporting format.
A stop condition. What result, by when, would cause them to recommend changing course. Anyone who cannot answer has not built a testable plan.
The questions that force a straight answer
- Which of your deliverables would change if we told you we did not care about AI search at all?
- What percentage of the proposed work is on our own site versus off it?
- Show me the last accuracy audit you ran for a client. What did it find?
- What is your measurement unit — keywords, prompts, or both?
- What has AI automation done to your internal cost per deliverable, and how is that reflected in our price?
- What is your verification step on AI-assisted output, and who performs it?
What to do in the next ninety days
First thirty days. Establish which model you are already buying. Ask your incumbent the five questions above — the answers are diagnostic, and this costs nothing. In parallel, run a manual prompt check to see whether you have a Model B problem at all.
Days thirty to ninety. If you have a gap, close the cheapest part first: the accuracy audit. Errors are fast to fix and expensive to leave. Then address retrievability on commercially important pages.
Beyond ninety days. Corroboration work, which runs on other people's publishing schedules.
One standard: the date a change ships is not the date its effect can be read. Allow three measurement cycles before concluding anything, and treat any single-week movement as noise. Nobody can guarantee a ranking, an indexing outcome or an AI citation — what can be committed to is a hypothesis, the evidence that would settle it, and the date you will know.
Where this goes next
The two models are converging, and within two years the distinction in this article will probably read as a period detail. Production tooling is becoming table stakes, and AI-surface optimisation is being absorbed into what "SEO" means. The agencies that survive that convergence will be the ones that can demonstrate both a verification standard on production and a measurement practice on prompts.
What will not converge is the pricing question. A service whose cost base has fallen and whose price has not is a temporary arrangement, and buyers are getting better at asking. The agencies worth keeping will volunteer that conversation before you raise it.
Take the five questions into your next agency review — including ours.
If you want to know which model you are currently buying and what the gap is costing, that is what an AI SEO and AEO engagement opens with: the accuracy audit, the prompt baseline, and the source-level gaps your rankings report cannot show you. WKND has been running search programmes since 2018 across more than 400 brands, with 20+ years of collective DTC and B2B experience behind the work.
