Quick answer

A credible GEO offer doesn't sell "a spot in ChatGPT." It sells an auditable protocol: map the buying questions, measure mentions and citations with repeats, fix the SEO obstacles, produce original evidence, and track results without confusing presence, click, and revenue. The quote depends on the number of markets, engines, prompts, repeats, pages, and validations. The agency guarantees its deliverables and its method, never a third party system's answer.

Key takeaways

  • GEO, AEO, and SEO share a foundation; don't bill the same fixes twice under a new name.
  • A measurement with no prompt, engine, country, date, and denominator isn't comparable.
  • Citations, mentions, visits, and conversions are four different outcomes.
  • The price should follow the work, the risk, and the coverage, not a promise of citation.
  • A good offer includes exclusions, frequency, re-baseline conditions, and the right to stop an inconclusive test.

Definition: what the agency actually sells

GEO, or generative engine optimization, here means the work aimed at improving a brand's and its content's ability to be correctly understood, found, mentioned, or cited in generated answers. That's a working definition, not a unified ranking system.

Google explains that its AI features can use a query fan-out technique: the system launches several related searches across subtopics and sources. Google also specifies that SEO fundamentals remain relevant and that no "special AI" file, markup, or length is required. The offer must therefore start with accessibility, quality, entities, sources, and measurement, not with hacks presented as certified.

The agency controls: the research, the protocol, the fixes, the content, the documentation, the tests, and the reporting. It doesn't control: the models, their corpora, their sources, the display frequency, the interfaces, personalization, or vendor changes.

The three offer levels

Level Suits Deliverables Measurement Main exclusions
Diagnosis A brand unaware of its AI visibility Intent map, multi-engine baseline, technical and entity audit, priorities One initial window with repeats No production and no guarantee of progress
Activation A brand with pages and experts available Diagnosis plus briefs, priority fixes, an original asset, source control Baseline then scheduled re-reading Press relations, heavy development, publishing with no validation
Operations An agency or multi-market company A versioned prompt portfolio, continuous production, tracking, incidents, QBR A time series per engine and market Citation guarantee, automatic causal attribution, markets not agreed

Each level must reuse the previous work. A client moving from diagnosis to operations doesn't pay a second time to recreate the same taxonomy; they pay to update and operate it.

Procedure: building the offer in nine steps

1. Choose a business outcome, not a GEO score

Start with the journey: category discovery, comparison, verification, selection, support, or reputation. A B2B vendor may target shortlist questions; a publisher may target citations on an area of expertise. Write down what would be useful even with no click.

2. Define the question population

Combine Search Console data, SERP research, sales interviews, tickets, reviews, and public questions. Classify each prompt by intent, persona, market, language, and stage. Threads where people ask whether AI SEO is real or just a gimmick show an objection to handle; they don't prove its volume.

3. Freeze a reproducible protocol

Record: engine and visible version, date and time, country, language, sign-in state, exact prompt, context provided, number of repeats, and capture policy. Since answers vary, a single run must not become a monthly statistic.

4. Separate the metrics

  • Mention rate = answers mentioning the brand ÷ answers tested;
  • Citation rate = answers containing a link attributable to the brand ÷ answers tested;
  • Intent coverage = intents with at least one eligible piece of content ÷ intents tracked;
  • Observed share of voice = brand mentions ÷ mentions of the defined competitive set;
  • Observable referral traffic = sessions identified from the tracked domains;
  • Influenced value = value under a stated rule, never automatically incremental.

Don't add these metrics into a magic score. A brand can be frequently mentioned but described inaccurately, or rarely cited on questions very close to purchase.

5. Audit the foundation before creating content

Check crawling, indexing, rendering, canonicals, permitted snippets, entity consistency, up-to-date information, and the ability to provide evidence. An inaccessible or contradictory page doesn't become "GEO-ready" thanks to an FAQ.

6. Look for information gain

Choose what the brand can publish that the existing pages don't hold: a dataset, a protocol, a calculator, a controlled example, method documentation, an important limitation, or the opinion of an identifiable expert. A generic synthesis of ten competitors isn't a source asset.

7. Create and check in batches

Produce a small set of pages or improvements. Check accuracy, sources, voice, rights, rendering, visible structured data, and cannibalization. Keep a control where reasonable. Don't create a page for every fan-out variation; Google warns that large scale production with no value can amount to abuse.

8. Measure using the planned window

Replay the same portfolio and record engine changes. Add the classic metrics, because a useful improvement can appear in Search or conversions without changing the AI answers. Conversely, an extra mention doesn't prove revenue.

9. Decide: maintain, deepen, or stop

Continue if the signal is reproducible and the value plausible. Revise the protocol if variance dominates. Stop a tactic if it requires unsourced claims, artificial mentions, community spam, or production with no demand.

Pricing: billing for coverage and risk

A defensible price starts from the cost of service:

Floor price = framing cost + measurement cost + production cost + QA cost + variance reserve + target margin.

For operations:

Measurement load = markets × engines × prompts × repeats × frequency × controlled unit cost.

The unit cost includes vendor, storage, anomaly review, and reporting. Add content work, development, original data, and press relations separately. Don't bill "per citation obtained": that encourages picking easy prompts, blurs the denominator, and promises an outcome outside your control.

Worked example: an offer for a B2B SaaS

A fictional example. A financial management SaaS sells in France and French-speaking Belgium. The agency selects 40 prompts: 10 category, 15 comparison, 10 risk, and 5 brand. It tests four engines with three repeats per quarter:

40 prompts × 4 engines × 3 repeats = 480 answers per window and per market configuration.

The baseline records 26 mentions and 9 citations out of 480 answers. Those values are 5.4% and 1.9% within this protocol only; they don't describe "all of ChatGPT" or "the whole market." The audit shows the comparison pages document neither method nor criteria, and that the pricing information is inconsistent between two pages.

The activation cycle fixes the consistency, publishes a calculator with visible assumptions, and turns two internal studies into reproducible methodologies. At retest, the agency will compare the same denominator, document interface changes, and also examine leads. Even if citations double, it will write "association observed after the changes," not "the pages caused the doubling" without a more robust design.

Clauses the quote must include

Clause Operational wording Risk avoided
Guaranteed deliverables List, format, date, and acceptance criteria Confusing work with a third party outcome
No guarantee No guarantee of mention, citation, ranking, or traffic An unverifiable promise
Baseline Protocol, denominator, engines, and window A score with no reference
Re-baseline Triggers: a major engine, market, or portfolio change An invalid comparison
Dependencies Experts, access, validation, development, data Delays wrongly attributed
Ownership Ownership of prompts, exports, content, and assets Lock-in or dispute
Security Minimum access, retention, and revocation Exposure of client data
Stop Spam, legal risk, missing sources, or excessive variance Continuing a harmful tactic

What the data proves and doesn't prove

In May 2026 Google announced a more continuous experience between AI Overviews and AI Mode. Its July 2026 documentation confirms fan-out, the importance of the fundamentals, and the absence of any mandatory special markup. That proves the position and behavior Google states; not a universal citation recipe.

The Pew study published in July 2025 observed 68,879 Google searches made in March 2025 by 900 American adults: users clicked a classic result in 8% of visits with an AI summary, versus 15% with no summary, and a source of the summary in 1%. The study is observational, American, and based on one month; it doesn't predict every sector or the 2026 experience.

Quora questions on the criteria for evaluating an AI optimization agency and Reddit on the best AI tracker for agencies signal demand. Quora and the GEO communities carry a lot of self-promotion: use the questions, not their vendor rankings, as evidence of need.

Common failures and stopping conditions

  • Renaming an SEO audit a "GEO audit" with no new measurement or deliverable.
  • Promising stable presence in a variable engine.
  • Testing only prompts that already contain the brand.
  • Changing the portfolio after the result without keeping the old denominator.
  • Presenting a mention without checking its accuracy or sentiment.
  • Creating fake reviews, citations, or community discussions: stop immediately.
  • Producing one page per subquestion with no demand and no unique value.
  • Selling an ROI calculated from mentions with no conversion data.
  • Continuing with a tracker whose results aren't reproducible and aren't exportable.

Reusable asset: the SEOryon GEO offer sheet

Use the AI visibility protocol to version prompts, engines, repeats, mentions, and citations. The rows provided illustrate the format; they aren't client results.

The sheet fits on two pages: objective, market, engines, prompt portfolio, repeats, separated metrics, baseline, deliverables, exclusions, dependencies, validation level, price, frequency, data ownership, re-baseline condition, and stopping condition.

Add a reproducibility index:

Reproducibility = prompts whose status is identical across at least two repeats ÷ prompts tested.

Define "status" before the test: not mentioned, mentioned, cited with a URL, cited with no URL. The index measures your protocol's stability, not an engine's absolute quality.

How SEOryon fits in

SEOryon can track mentions and citations in ChatGPT, Perplexity, Gemini, and Claude, analyze SERPs and questions, recommend or write content to a brand voice, check cannibalization, read Search Console and Analytics, and publish to several CMSs. The choice between semi-autopilot and autopilot modes lets you adjust the control.

These functions support collection and execution. They guarantee no citation, don't prove a causal link, and replace neither a versioned protocol nor expert validation.

Measurable exercise

Build a diagnosis offer for a real brand without executing it. Deliver twenty prompts, four intents, at least two engines, three planned repeats, a deliverables and exclusions matrix, and a quote based on the load. Success if:

  1. every KPI has a denominator;
  2. the quote promises no third party outcome;
  3. SEO and GEO aren't billed twice;
  4. a buyer can reproduce the protocol;
  5. a stopping condition and a re-baseline condition are written down.

FAQ

How much should you charge for GEO work?

Calculate the framing, the measurement coverage, the repeats, the production, the control, the tools, the variance, and your margin. A flat fee is only defensible if those assumptions and the overage terms are stated.

Can you guarantee a citation in ChatGPT or Google AI Mode?

No. Guarantee the deliverables, the protocol, and the controls. The answers, models, sources, and interfaces belong to variable third party systems.

What's the difference between an SEO and a GEO offer?

The technical and quality foundation overlaps. GEO adds a map of questions, a multi-engine protocol, mention and citation metrics, and work on sourceworthiness. The detailed distinction is in SEO, GEO, and AEO.

What's the best GEO KPI?

There is no single KPI. Measure mention, citation, accuracy, share of voice, traffic, and value separately, then choose the one matching the business decision.

How many prompts should you track?

Enough to cover the important intents, but few enough to allow repeats and control. Start with a stratified portfolio and expand only once the measurement is stable.

References

Method and update note

The example's volumes are fictional; the Pew statistics keep their population, period, geography, and limitation. Reviewed 16 July 2026, translated and edited 22 July 2026. Revisit this guide on a major change to Google's features, the Search Console reports, the engines SEOryon tracks, or the offer's commercial terms.