A generative engine optimization agency works on whether AI systems retrieve, cite and describe your business correctly when someone asks a question in your category. In practice that means content that is easy to extract and quote, corroboration on sources those systems already trust, consistent facts about your business wherever they appear, and a measurement layer that tells you when any of it moves. It overlaps heavily with SEO, and the honest ones will tell you that before you ask.
The category is eighteen months old, the pricing is unstandardised, and a meaningful share of what is being sold as GEO is repackaged SEO with a monitoring subscription attached. This article is the vetting process: what the work consists of, what moves the price, which claims fail against primary documentation, and the strongest argument that you should not hire anyone for this at all.
What the work actually consists of
Strip the vocabulary away and four things are being worked on.
Retrievability
Whether a machine can get your content, parse it, and pull a clean statement out of it. This is mostly unglamorous technical work: crawlable pages, important content in text rather than locked inside images or rendered only after interaction, clear headings, and claims stated plainly enough to be lifted without distortion.
Google's own position is worth holding onto here, because it deflates a great deal of GEO marketing. Google states that "the best practices for SEO remain relevant for AI features in Google Search" and that "there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary." The eligibility bar it describes is ordinary: a page must be indexed and eligible to appear with a snippet.
An agency telling you that AI visibility requires a fundamentally different technical discipline is, on Google's surfaces, contradicting Google.
Source corroboration
AI systems assemble answers from multiple places. If your claim appears only on your own site, it is a marketing assertion. If it appears on your site and is corroborated by sources those systems already draw on, it behaves more like a fact. A real GEO engagement therefore includes off-site work — the sources, directories, comparison pages and third-party coverage in your category.
Semrush's AI Visibility Toolkit documents this directly as Source Opportunities: sources frequently cited in AI answers that mention competitors but not you. Ahrefs' Brand Radar tracks the same shape through Citations — how often AI links to your pages or domains as a source — and extends it to YouTube, TikTok and Reddit. Both vendors, independently, have built product around the premise that off-site corroboration is where the work lives.
Entity and claim consistency
Whether your business is described the same way everywhere it appears. Conflicting service descriptions, outdated pricing, an old positioning statement still sitting on a directory profile — these produce confidently wrong AI answers, and they are among the cheapest problems to fix once found.
The scope is wider than most teams expect. It includes your own site, but also your LinkedIn company page, your Crunchbase and G2 profiles, directory listings, partner pages that describe you, conference speaker bios, old press releases, and any review platform in your category. HubSpot makes the same point in its own AEO guidance: a strong strategy "goes beyond your own site to include your presence on LinkedIn, Reddit, YouTube, third-party blogs, affiliate partners, and review sites, covering everywhere answer engines look."
The failure pattern is predictable. A company repositions, updates its homepage, and leaves fourteen other descriptions of the old positioning live across the internet. The systems weigh corroboration, and thirteen sources agreeing with each other will outweigh one disagreeing — even when the one is you.
Two practical examples, because this is vaguer than it sounds. A pricing page that says "flexible pricing tailored to your needs" cannot be quoted by anything. A pricing page that says "engagements start at X and scale with Y" can. A service page whose differentiator is buried in the fourth paragraph of a case study will lose to a competitor who states it in a sentence under a heading. None of that is an AI technique. It is writing that can be extracted, and it happens to be what the systems reward.
Measurement
Whether anyone can tell if the work did anything. This is the component most likely to be missing, and the one most worth interrogating in a pitch. The mechanics of doing it properly are covered in our piece on tracking AI visibility.
A measurement plan that deserves the name specifies four things before any work begins: which prompts are being tracked, on which engines, how often, and what the reading was at baseline. Without the baseline, every subsequent number is unfalsifiable. An agency that starts work before establishing one has removed your ability to hold them to anything — and in our experience that omission is far more common than deliberate overclaiming.
GEO or SEO with a new label: five tells
Being blunt, because this is the question most buyers are actually asking.
One. Ask what percentage of the proposed work would also be recommended for a client with no interest in AI search. If the honest answer is most of it, you are buying SEO. That is not necessarily bad — it may be exactly what you need — but you should not pay a premium for the prefix.
Two. Ask which platforms they measure and how. "AI visibility" spans engines with different behaviour. Ahrefs documents coverage across AI Overviews, AI Mode, ChatGPT, Perplexity, Gemini and Copilot; Semrush's documented coverage is ChatGPT, Gemini and AI Overviews. A proposal that says "AI search" without naming engines has not been scoped.
Three. Ask what they will change off-site. If the entire plan is on your own domain, it is missing the corroboration half.
Four. Ask what they would do if you already ranked well. A GEO service that is only useful to sites with weak SEO foundations is a foundations service.
Five. Ask what result would make them tell you to stop. Anyone who cannot describe a losing outcome has not thought about measurement.
A sixth question is worth adding if you are comparing proposals rather than evaluating one: ask each agency what they would do differently from the others on your shortlist. Vendors who understand their own positioning can answer this crisply. Vendors selling a template cannot, because they have not thought about where their approach sits in the market.
What this looks like at different stages of maturity
The right engagement depends heavily on where you are starting, and proposals that ignore this are templates.
You have strong SEO and good brand coverage. Your foundations are already doing most of the work, and the marginal gain is in corroboration and accuracy — making sure the substantial coverage you already have says the right things, and filling the source gaps where competitors appear and you do not. This is the cheapest starting position and the one where GEO-specific work is most clearly additive rather than remedial.
You have strong SEO and thin brand coverage. Common in performance-led businesses that have invested in acquisition but not in earned media. Your pages are retrievable and your claims are uncorroborated, so systems can find you but have little external reason to trust you. The work is off-site and it is slow, because it depends on other people publishing.
You have weak SEO and strong brand coverage. Often the case for established businesses with an ageing site. Systems know who you are from everywhere else, and your own pages are not competitive as sources. The fastest wins are on-site, and they are ordinary technical and content SEO wearing a different label. An honest agency will tell you this and price it accordingly.
You are early on both. The sequence is foundations first. Google states plainly that "the best practices for SEO remain relevant for AI features in Google Search" and that eligibility requires a page to be indexed and snippet-eligible. There is no route that skips this, and any proposal suggesting otherwise is selling a shortcut that does not exist.
The verification gap
This is where a vetting process earns its keep. Three claims appear repeatedly in this category and can be checked against primary documentation in about ten minutes.
"We'll add an llms.txt file so AI can read your site." — SETTLED, and it fails. Google's documentation states: "You don't need to create new machine readable files, AI text files, or markup to appear in these features." There is no ambiguity in that sentence. We do not ship these files and we do not quote for them.
"Our proprietary AI schema gets you cited." — SETTLED, and it fails. The same page: "There's also no special schema.org structured data that you need to add." Ordinary structured data remains worthwhile for other reasons, and Google asks that it match the visible text on the page. But it is not an AI citation lever, and no agency has a proprietary version of an open standard.
"We'll guarantee AI Overview placement." — SETTLED, and it fails. Google: "Just because a page meets all requirements, best practices, and complies with the policies, doesn't mean that Google will crawl, index, or serve its content. Indexing and serving isn't guaranteed." Anyone guaranteeing placement is guaranteeing something the platform explicitly declines to guarantee.
Where this gets genuinely interesting is the question those three claims obscure. Google publishes guidance. ChatGPT, Perplexity and Copilot largely do not. Whether structured signals matter more on systems that have not documented their behaviour is contested — and an agency that distinguishes the two cases is doing the thinking. One that applies Google's guidance to every engine, or ignores it entirely, is not.
What a real scope of work contains
If you take one artifact from this article into a vendor conversation, make it this. A GEO scope of work that can be held to account names all seven of these. Most proposals name two or three.
One — the engine list. Which systems are in scope. Not "AI search." Named engines, with a note on which are measured and which are only optimised for.
Two — the baseline. Current state on those engines, measured before work starts, using a documented prompt set.
Three — the prompt set itself. Written down, versioned, and yours. If the prompt set belongs to the agency, your measurement history dies when the engagement does.
Three and a half — the measurement baseline. Separate from the prompt set: the actual reading at day zero, across every engine in scope, recorded before a single change ships. Without it, every subsequent number is unfalsifiable and nobody can be held to anything.
Four — the on-site work. Specific pages, specific changes, with an honest note on which of them are ordinary SEO improvements. A proposal that cannot separate these is either unsophisticated or hoping you will not ask.
Five — the off-site work. Named source categories and a method for earning placement in them. "Digital PR" is not a method. "These six comparison sites and review platforms are cited in your category's answers and we will pursue them in this order" is.
Six — the reporting model. Cadence, format, and what gets reported when the number goes down.
Seven — the stop condition. What result, by when, would cause the agency to recommend changing course. Anyone who cannot answer this has not built a testable plan.
What it costs and what moves the number
There is no standard rate card, and anyone quoting one as an industry benchmark has invented it. What can be said is what drives the number.
Off-site work is the main cost driver. On-site retrievability improvements are bounded — there are only so many pages. Source corroboration is open-ended and involves other people's publishing timelines.
Measurement is a real line item, and the tooling is public. Semrush's AI Visibility Toolkit is $99/month, and at that tier includes 25 prompts for tracking. Ahrefs' Brand Radar runs $50/month for custom prompts or $199/month for its AI Visibility Index at 83 prompts per day. If a proposal includes monitoring, ask whether that is a pass-through cost or margin.
Scope of engines multiplies everything. Six platforms is not twice the work of three.
Existing foundations change the shape completely. A site with unresolved technical problems will spend the first months on SEO regardless of the label on the invoice.
Three engagement shapes are common, and they price very differently:
Audit and roadmap. A fixed-scope diagnostic producing a baseline, a gap analysis and a prioritised plan. Useful when you do not yet know whether you have a problem. It should not roll automatically into a retainer.
Retainer. Ongoing content, corroboration and measurement work. Appropriate once a baseline exists and there is identified work to do. The risk is paying indefinitely for monitoring dressed as optimisation — which is why the stop condition matters.
Project. A defined body of work with a defined end, most often a content and structure overhaul on a specific set of pages. Often the best value if your gap is on-site rather than off-site.
The shape should follow the diagnosis. An agency proposing a retainer before establishing a baseline is proposing the shape that suits them.
The GEO vendor scorecard
Ten factors, rated 1 to 5, where 1 is weak and 5 is strong. This is the same rubric we use internally on engagements, which is the only reason it is worth anything.
| # | Factor | What a 5 looks like |
|---|---|---|
| 1 | Scope clarity | Named engines, named deliverables, named exclusions |
| 2 | Off-site plan | Specific sources and a method for earning them |
| 3 | Measurement | A defined prompt set, cadence and reporting model before work starts |
| 4 | Claim accuracy | No llms.txt, no proprietary schema, no placement guarantees |
| 5 | Baseline | They measure where you are before quoting |
| 6 | Foundations honesty | They tell you if the real problem is technical SEO |
| 7 | Reporting | Shows losses as well as gains |
| 8 | Ownership | You keep the prompt sets, documentation and assets |
| 9 | Exit terms | You can leave without losing the measurement history |
| 10 | Stop condition | They can describe what failure looks like |
Score = 100 × sum(rating − 1) / 40.
Worked through: a proposal scoring 4 on scope clarity, 2 on off-site plan, 3 on measurement, 5 on claim accuracy, 2 on baseline, 4 on foundations honesty, 3 on reporting, 3 on ownership, 2 on exit terms and 1 on stop condition sums to 29. Subtract 1 per factor across ten factors, giving 19. 100 × 19 / 40 = 47.5.
Forty-seven is not a rejection. It is a list of five specific conversations to have before signing — off-site plan, baseline, exit terms, ownership and stop condition — and a vendor who improves three of them has materially improved the engagement. The score's job is to make the gaps visible, not to produce a verdict.
For any factor you genuinely cannot assess, do not substitute a middle rating. Score the range — the lower bound using 1 and the upper using 5 — and treat the gap as the thing to resolve before signing. A vendor you cannot score is a vendor you have not finished evaluating.
What the numbers don't agree on
There is an open question underneath every GEO proposal, and most of them step around it: what should the engagement actually be measured on? Two credible answers exist, they produce different scorecards, and no platform has settled it.
Position A — citation share. Measure how often you are cited or mentioned relative to competitors. Both major platforms have built product on this. Semrush reports an AI Visibility Score, 0–100, which its documentation says "reflects how often your brand is mentioned across AI platforms," and surfaces Source Opportunities — sources cited in your category that mention competitors but not you. Ahrefs reports AI Share of Voice, the percentage of responses within a topic set mentioning or citing you versus competitors, plus Estimated Impressions weighted by the real search volume behind each prompt.
The logic is sound: share of voice is comparable, trackable, and competitive. It answers "are we present?"
Position B — representational accuracy. Measure whether what these systems say about you is correct. Not how often you appear, but whether the appearance helps. A brand can lift its citation share while being consistently described with the wrong services, outdated pricing, or a positioning it abandoned two years ago.
The logic here is also sound, and it is closer to the commercial outcome: a confident wrong answer in front of a buyer is worse than no answer. It is also harder to quantify, which is precisely why most tools lead with Position A. CONTESTED — and the fact that both vendors report mentions and citations as separate metrics rather than merging them tells you the industry has not resolved it either.
Our read. In the first two quarters of an engagement, weight accuracy first. Errors are the cheapest category of problem to fix, the fastest to move, and the most directly damaging while they persist — and in our experience they are more common than clients expect, because nobody has looked. Once the accuracy column is clean, citation share becomes the primary KPI, because at that point more visibility is unambiguously good rather than amplification of something wrong.
Weight volume when you can. Ahrefs' approach of weighting by the search demand behind each prompt is the more decision-useful model, because appearing in a prompt nobody asks is not worth the same as appearing in one thousands do. A flat prompt set treats both equally, and that is a real limitation to know about rather than a reason to avoid tracking.
What would change our position: platform-published data connecting citation frequency to downstream conversion. Nobody has it yet. If a platform releases it and citation share correlates cleanly with revenue, Position A becomes the correct primary metric from day one and we would say so.
A proposal that names which of these it optimises for — and can defend the choice — is a proposal from someone who has thought about the problem.
Questions that separate operators from resellers
- What would you change in the first thirty days, and what evidence would tell us it worked?
- Which engines are in scope, and which are not?
- What is your prompt set, and do we keep it if we leave?
- What is the largest problem with our current site that has nothing to do with AI?
- Which of your recommendations would you drop if we told you budget was halved?
Where this goes next
Two shifts are worth planning for over the next twelve to twenty-four months, and neither is a bigger retainer.
The measurement layer is going to move first-party. Google already reports AI-feature performance inside Search Console. Others will follow, because the alternative is ceding the narrative to third-party estimates. When that happens, the value of an external visibility score drops sharply and the value of knowing what to change with it does not. Engagements built around a dashboard will be exposed. Engagements built around content and corroboration will not.
The vocabulary is going to consolidate. GEO, AEO and LLMO describe overlapping work with unsettled boundaries, and at least one of those terms will not be in common use in two years. That is an argument for choosing a partner on the basis of what they actually change rather than which acronym they lead with — and for writing your own internal definitions down now, so your reporting survives the churn.
The brands that are visible in assistants in two years will mostly be the ones whose claims were easy to verify and easy to quote. That is a less exciting conclusion than most of this category's marketing, and it is the one the evidence keeps producing.
Run the scorecard on whoever is pitching you — including us. If you would rather we walked through it against your own site and current visibility, that is a GEO scoping call. We have been running search programmes since 2018 across more than 400 brands, and the pattern in this article is the one that keeps repeating.
