# Which KPIs can an agency promise in a GEO retainer?

> Agencies can promise defined GEO work, monitoring, reporting, and service levels, but not guaranteed placements in ChatGPT or Google AI Overviews.

Source: https://bestgeoplatforms.com/geo-retainer-kpis · Publisher: GEO Platforms for Agencies · Updated: 2026-11-15

## TL;DR
- Agencies can promise defined GEO work, monitoring, reporting, and service levels, but not guaranteed placements in ChatGPT or Google AI Overviews.[24]
- Mention rate, citation share, and share of voice can be contractual targets only when the formula, baseline, panel, timeframe, assumptions, and remedy are explicit.[4][27]
- Repeated observations are essential because AI answers vary, and small prompt samples can produce wide uncertainty.[4]
- Google Search Console now reports generative-AI impressions, while GA4 separately classifies recognized assistant traffic and keeps Google AI Overviews and AI Mode in Organic Search.[11]
- Monthly reports should show measured visibility and delivery; quarterly reviews should assess trend reliability, attribution limits, source quality, and commercial evidence.[2][16]

An agency can responsibly promise defined work, measurement coverage, reporting quality, and—if the baseline, formula, timeframe, assumptions, and remedy are explicit—targets for metrics such as mention rate, citation share, share of voice, and analytics-attributed AI sessions. It should not guarantee a specific organic placement in ChatGPT, Google AI Overviews, or another answer engine because eligibility does not create an entitlement to appear, and platforms do not guarantee individual results.[1][2]

The distinction matters in a GEO retainer. A client can hold an agency to a documented prompt panel, completed optimization work, collection frequency, reporting deadlines, and agreed data-quality standards. A client cannot reasonably treat one changing AI response as a guaranteed ranking position.

**Illustrative example:** an agency might commit to increasing the brand’s mention rate from an agreed baseline over a quarterly review window, provided the prompt set, engines, markets, repeat schedule, and calculation remain fixed. It should not promise that “the brand will appear in the top three ChatGPT recommendations for every target prompt.”

## What is a GEO KPI, and what can an agency promise?

A GEO KPI is a defined measurement of how a brand, its content, or its website appears in generative-search answers and related traffic or business data. Common measures include brand mention rate, owned-domain citation rate, citation share, share of voice, Google generative-AI impressions, AI-assistant sessions, and selected conversions.[3][2]

The safest contract structure separates three layers:

1. **Service obligations:** audits, prompt research, content updates, technical recommendations, monitoring, reporting, and agreed response times.
2. **Measured performance targets:** movement in a specified metric from a documented baseline over a defined period.
3. **Platform outcomes:** inclusion, citation, position, or recommendation in an individual AI answer.

The first layer is directly within the agency’s control. The second can be negotiated if the measurement design and dependencies are clear. The third depends on changing platforms, retrieval systems, source selection, user context, and answer generation, so it should not be sold as certain.[4][2]

Google states that no one can guarantee a number-one search ranking and warns buyers about supposed special relationships or priority submission.[5] Google’s AI-feature documentation similarly explains that pages must meet eligibility conditions but that indexing and serving are not guaranteed.[6]

### KPI table: what belongs in a GEO retainer?

| KPI | Definition | Can it be promised? | Why |
|---|---|---|---|
| Mention rate | Answers naming the brand divided by all valid tracked answers.[7] | Yes, as a defined target or reporting obligation | The agency can specify the panel, repeats, calculation, and review period, but cannot control each answer. |
| Owned-domain citation rate | Answers citing the brand’s domain divided by all valid tracked answers.[7] | Yes, with explicit scope | A citation is observable, but platform retrieval and citation behavior remain external dependencies.[8] |
| Citation share | The brand’s share of citations within the selected result set; the exact denominator varies by vendor.[9] | Yes, only after defining the vendor formula | “Citation share” is not universal: platforms count citations and denominators differently.[4] |
| Share of voice | The brand’s mentions divided by the relevant category or competitor mentions.[3][7] | Yes, as a bounded comparative target | Competitor activity and the selected competitor set can change the result.[9] |
| Visibility rate | Depending on the tool, answers naming the brand, citing its domain, or both.[7] | Only with a written formula | Vendor labels differ, and one answer may contain both a mention and a citation.[7][4] |
| Google generative-AI impressions | Site-link impressions from Google AI Overviews and AI Mode in Search Console.[10] | Yes, as a reporting KPI; cautiously as a target | Google reports impressions, but the documentation does not establish a complete AI-specific click, position, or conversion funnel. |
| AI-assistant sessions | Analytics sessions classified through recognized assistant traffic sources, such as ChatGPT, Gemini, or Copilot.[11][12] | Yes, as an analytics target with settings disclosed | GA4’s channel rules differ from Google Search reporting, and Google AI Overviews and AI Mode remain in Organic Search.[8][12] |
| AI-attributed key events | Conversions or other events associated with a specified AI traffic source and attribution model.[13] | Yes, as attributed activity—not guaranteed causation | Attribution models, direct traffic, modeled events, and multiple marketing influences limit causal certainty.[14] |
| ChatGPT placement | The brand appearing in a particular ChatGPT response or position | No, not as a guaranteed organic outcome | OpenAI documents referral tracking but does not document a guaranteed placement or publisher ranking mechanism in the cited source.[15] |
| Google AI Overview placement | The brand or page appearing in a particular AI Overview | No, not as a guaranteed organic outcome | Google says eligibility does not guarantee serving.[6] |
| Exact answer wording or position | A specified wording, order, or recommendation slot in an answer | No | Generative answers vary across runs and may change with context, sources, model updates, and location.[4] |

The contract should identify the **tool, report, field, numerator, denominator, engines, markets, languages, prompt version, repeat schedule, valid-observation rule, and remedy**. A promise such as “improve GEO visibility” is too vague to evaluate.

## How should an agency set GEO retainer targets from a baseline?

AMEC recommends agreeing the audiences, markets, languages, competitors, models, questions, narratives, and testing conditions before measuring change.[2] Baselines should begin when monitoring begins, with trends built prospectively rather than reconstructed from incomparable historical snapshots.[2]

Use this five-step method:

1. **Define the commercial scope.** Select the markets, languages, customer groups, competitors, engines, answer surfaces, and prompt themes that matter to the client. Include both branded and unbranded prompts where appropriate.[2] Treat each answer surface as its own scope: [Finseo's own comparison of Google AI Mode and AI Overviews](https://www.finseo.ai/blog/ai-mode-vs-ai-overviews) found the two share only 13.7% of cited URLs on the same query.
2. **Version the prompt panel.** Record each prompt, its intent, engine, market, language, collection date, repeat count, and inclusion or exclusion rule. Keep a stable core panel even when new prompts are added.
3. **Run repeated baseline observations.** Collect enough valid answers to estimate the metric rather than relying on one screenshot or one run. Report the numerator, denominator, failed collections, and exclusions.
4. **Set a target that matches measurement precision.** Choose an absolute percentage-point change, relative change, range, or minimum threshold only after reviewing volatility and uncertainty. Small samples cannot support precise promises.[4]
5. **Write the review and remedy rules.** State the measurement window, data-finalization date, platform dependencies, methodology-change process, client obligations, and any credit, fee adjustment, or corrective action.

There is no universal mandatory GEO sample size in the cited AMEC guidance.[2][16] SparkToro’s research involved 600 volunteers, 12 prompts, three tools, and 2,961 runs; its separate headphone analysis used 142 prompts and 994 responses.[17] That supports repeated aggregate measurement, not a universal prompt-count rule.

A 2026 preprint based on 4,320 observations found repeat instability ranging from 1.4% to 40.3% of prompt-platform cells. At 12 prompts, the least precise platform’s 95% citation-rate interval spanned 53 percentage points, while the smallest resolvable differences ranged from 25 to 36 points.[18] Those findings make a five-point guaranteed improvement especially difficult to defend when the panel is small.

## Why do answer variance and sample size change what can be promised?

AI answers can change even when the prompt appears identical. Model updates, retrieval results, location, language, timing, source availability, personalization, and tool behavior can all affect the observed answer.[2][4] In [Finseo's own study of about 1.4 million German-language (DACH) AI answers](https://www.finseo.ai/blog/ai-citation-stability), only 38.4% of URLs cited by ChatGPT were still cited the next day for the same prompt, and a single day of ChatGPT visibility missed the following fortnight by 16 points on average.


How should an agency explain answer variance and sample size to a client?

A single answer is an observation, not a stable ranking. Repeating the same prompt helps estimate how often an event occurs, such as whether the brand is named or its domain is cited, but repeated answers may still vary substantially.[8]

The SOW should define the prompt panel, engines, markets, repeat schedule, valid observations, failed collections, and confidence or uncertainty reporting. It should also state whether a target applies to the full panel, a segment, or a rolling average. SparkToro’s research supports measuring aggregate visibility across repeated prompts, while the scientific preprint shows that small samples can produce wide uncertainty.[4]

Do not describe a percentage as more precise than the design supports. If the sample cannot reliably distinguish a five-point change from ordinary variation, use a broader range, increase repetitions, or make the KPI observational rather than performance-linked.



The practical implication is to promise **measurement quality and transparency** before promising a narrow percentage-point movement. If the client wants a performance-linked fee, the parties should agree what degree of movement is detectable and how missing or unstable observations are handled.

## How should mention rate, citation share, and share of voice be defined?

Metric names are not standardized across GEO platforms. Finseo defines mention rate as answers naming the brand divided by all tracked answers, citation rate as answers citing the domain divided by all tracked answers, and visibility as answers naming the brand or citing its domain.[7]

Peec describes visibility as the percentage of AI responses in which the brand appears, while its citation share measures the portion of inline citations attributed to a domain or URL.[19] Profound’s Answer Engine Insights overview uses a visibility denominator based on responses containing at least one brand.[3][9] Semrush’s AI Visibility Score is a composite of topic coverage and mention consistency rather than a simple percentage of answers.[1]

That means a SOW should avoid undefined phrases such as:

- “Increase AI visibility by 20%.”
- “Own 30% of citations.”
- “Reach first place in AI search.”
- “Appear in most ChatGPT answers.”

A stronger definition would be:

> **Illustrative KPI definition:** “Mention rate means the number of valid collected answers that name the client brand, divided by the number of valid collected answers in the agreed panel. The KPI will be segmented by engine and market and reported with numerator, denominator, collection dates, prompt-panel version, and failed-collection count.”

For citation share, specify whether the denominator is all citations, citations among selected competitors, citation occurrences, or answers containing citations. For share of voice, identify the competitor set and whether the calculation uses all detected brands or only named tracked competitors.[7]

Mention rate and citation rate should remain separate. A brand can be named without its website being cited, and a website can be cited without the brand being named.[19] Neither metric should be presented as a substitute for the other.

## Can an agency guarantee AI referral sessions or revenue?

An agency can agree to report AI-attributed sessions and selected key events, and it may negotiate a target if the analytics configuration, traffic sources, attribution model, and review window are explicit. It should not describe those figures as all traffic or revenue caused by GEO.[20]

GA4 introduced an AI Assistant channel in May 2026, with recognized assistant referrers receiving the `ai-assistant` medium and `(ai-assistant)` campaign.[11] Google AI Overviews and AI Mode remain within Organic Search rather than the AI Assistant channel.[8] OpenAI’s publisher guidance states that ChatGPT adds `utm_source=chatgpt.com` to referral URLs, allowing publishers to analyze incoming traffic.[15]

The reporting streams therefore have different scopes:

- **Search Console:** Google generative-AI impressions, with pages, countries, devices, and dates in the documented report.[10][21]
- **GA4:** recognized assistant sessions and events under the relevant channel and attribution settings.[11]
- **ChatGPT referral analytics:** visits carrying the documented ChatGPT UTM signal.[15]
- **Business systems:** conversions, revenue, pipeline, or orders joined through the client’s analytics and CRM setup.

GA4 session-scoped acquisition and event-scoped attribution use different rules.[22] Direct traffic may lack referral information, and modeled key events can estimate events that cannot be directly observed.[14] The contract should therefore use language such as **“analytics-attributed AI sessions”** or **“AI-source-associated key events,”** not “revenue generated by GEO.”

A performance fee can still be appropriate if it uses a defined attributed metric, includes a data-finalization delay, and acknowledges other influences. The ANA/4A’s guidance says incentive metrics should align with business goals affected by agency services and should be within the agency’s control, with transparent baseline-setting before implementation.[23]

## What should the monthly report contain?

The monthly report should answer: **What was measured, what changed, what was delivered, and how reliable is the observation?**

Include:

- Completed work against the agreed delivery plan.
- Approval delays, access issues, missed deadlines, and remedies.
- Prompt-panel version, engines, markets, languages, repetitions, valid observations, and failed collections.
- Mention rate with numerator and denominator.
- Citation rate and citation share as separate measures.
- Share of voice with the competitor set and formula.
- Cited domains and pages, including changes in source mix.
- Search Console generative-AI impressions.
- GA4 AI Assistant sessions and selected events.
- ChatGPT-source sessions where the UTM signal is available.
- Incorrect, outdated, harmful, or reputationally risky representations.
- Methodology changes and platform outages.

AMEC recommends a governed query library, documented testing conditions, saved outputs, source-quality review, and a risk log.[16] It also treats source-citation data as a time-sensitive snapshot and recommends prospective trend-building rather than unsupported claims about total impact.[2]

The monthly report should not collapse every number into one “GEO score.” A composite can be useful for internal navigation, but payment and client decisions should rely on clearly defined underlying measures.

## What belongs in the quarterly GEO review?

The quarterly review should evaluate whether the measurement system and strategy remain fit for purpose, not simply repeat the monthly dashboard.

Review:

1. Whether the prompt panel still represents the client’s priority audiences and commercial questions.
2. Whether observed changes exceed expected answer variation.
3. Whether competitor movement or a vendor methodology change affected the result.
4. Whether cited sources are accurate, current, useful, and credible.
5. Whether content and technical work improved the source environment.
6. Whether AI-attributed sessions or key events show a consistent trend.
7. Whether qualified pipeline, conversions, or revenue provide supporting commercial evidence.
8. Whether the KPI target, scope, resources, or remedy should be renegotiated.

AMEC distinguishes observed AI outputs from outcomes and impact.[16][2] A higher mention rate can indicate improved visibility in the tested panel, but it does not prove increased awareness, trust, consideration, or revenue.

Google’s generative-AI Search Console report provides an important new exposure signal: Google announced it on June 3, 2026, and later reported worldwide rollout on August 31.[10] The documented report focuses on impressions and does not establish a complete AI-specific click-and-conversion funnel.[4] Keep those impressions separate from analytics-attributed assistant sessions and from commercial outcomes.

## What should a GEO SLA say?

A GEO SLA should separate delivery commitments from performance targets and platform dependencies.

> **Illustrative SOW wording—not legal advice:**  
> “The Agency will collect and report [metric] using [tool and configuration], [prompt-panel version], [engines], [markets/languages], and [repeat schedule]. Each report will disclose the numerator, denominator, valid observations, failed collections, exclusions, and methodology changes.”
>
> “The parties agree a target of [value or change] from baseline [value and observation window], assessed over [review window]. Any performance-linked fee, credit, or other remedy is defined in [schedule]. This target is not a representation of guaranteed inclusion, citation, wording, or position in any individual AI response.”
>
> “Delivery dates, response times, client approval obligations, access requirements, and remedies for missed service levels are specified separately. Material changes to platform availability or measurement methodology trigger the agreed change-control process.”

This structure follows contract guidance that performance commitments should specify the metric, baseline, timeframe, assumptions, dependencies, and remedy.[24] It also reflects the practical distinction between holding an agency to a list of deliverables and holding it to a ranking or platform result.[20][23]

A defensible GEO SLA might guarantee:

- A monthly report delivered by an agreed business day.
- A fixed monitoring panel collected at an agreed frequency.
- A stated number of audits, content briefs, updates, or technical reviews.
- Documented response times for client requests.
- A defined process for correcting measurement failures.
- A target for a metric, if the baseline and calculation are fixed.
- A service credit or corrective action when the agency misses an agreed deliverable.

It should not guarantee:

- A specific ChatGPT recommendation.
- A fixed Google AI Overview placement.
- A permanent citation.
- A precise answer wording.
- A universal increase in AI visibility across every prompt and market.
- Revenue or pipeline causally attributable only to GEO.

## Which KPIs should agencies use for different client goals?

For **visibility**, use mention rate, citation rate, citation share, and share of voice, each with a written formula.[25]

For **source quality**, report the pages and domains cited, the types of sources appearing, and whether the client’s information is accurate, current, and useful.[16]

For **Google exposure**, use Search Console’s generative-AI impressions report rather than claiming that every organic visit came from an AI Overview or AI Mode result.[10]

For **assistant traffic**, use GA4’s AI Assistant channel and the documented ChatGPT referral signal, while keeping Google AI search traffic in Organic Search.[26]

For **commercial contribution**, report attributed sessions, key events, pipeline, or revenue with the attribution settings and data-finalization rules disclosed.[22][14]

For **agency accountability**, use completed deliverables, reporting timeliness, collection coverage, issue resolution, and client approval dependencies.[24][23]

The best GEO retainer does not promise less; it promises precisely what the agency can observe, influence, and remedy. It uses visibility metrics as evidence, traffic data as a separate evidence stream, and business outcomes as a broader quarterly question.

## FAQ
### Can a GEO agency guarantee ChatGPT placement?
No. An agency can optimize content, monitor a defined prompt panel, and report observed inclusion or citation rates, but it should not guarantee a specific ChatGPT placement or answer position. OpenAI’s cited publisher guidance documents referral tracking rather than a guaranteed organic placement mechanism.[16][15]

### Can an agency promise a higher GEO mention rate?
It can agree to a mention-rate target if the contract defines the prompt panel, engines, markets, repeats, baseline, timeframe, calculation, and remedy. The target should be presented as a measured performance objective, not as a guarantee that every individual answer will mention the brand.[27]

### What is the difference between mention rate and citation rate?
Mention rate measures answers that name the brand, while citation rate measures answers that cite the brand’s domain.[7] A response can name a brand without citing its website, or cite a website without naming the brand, so the metrics should not be combined without an explicit formula.[7]

### How many prompts are needed for a reliable GEO KPI?
There is no universal prompt count established by the cited AMEC guidance.[4][16] The required sample depends on the target precision, expected variation, number of engines and markets, and whether the agency uses repeated observations; research shows that small samples can produce very wide uncertainty.[4]

### Can a GEO retainer guarantee AI-generated revenue?
It should not guarantee revenue caused solely by GEO. An agency can report analytics-attributed AI sessions and key events under specified attribution settings, but multiple channels, direct traffic, modeled events, and other business influences limit causal interpretation.[4][22]

### Should GEO KPIs be reviewed monthly or quarterly?
Operational metrics and delivery obligations belong in the monthly report, including coverage, mentions, citations, traffic signals, and completed work. A quarterly review is better for judging trend stability, source quality, commercial evidence, target suitability, and changes to scope or strategy.[27][23]

## Sources
[1] Where does the data in Semrush’s AI Visibility Toolkit come from? — https://www.semrush.com/kb/1607-semrush-ai-visibility-data
[2] amec conflict.inddc — https://amecorg.com/wp-content/uploads/2026/05/26-05-15-A-Practitioners-Guide-to-GEO-Measurement.pdf
[3] About Answer Engine Insights (AEI) — https://docs.tryprofound.com/answer-engine-insights/about
[4] Measuring Brand and Source Discovery under Repeated LLM Queries:A Finite-Sample Audit — https://arxiv.org/html/2609.05059
[5] Do You Need an SEO? Tips for Hiring an SEO | Google Search Central  |  Documentation  |  Google for Developers — https://developers.google.com/search/docs/fundamentals/do-i-need-seo
[6] AI Features and Your Website | Google Search Central  |  Documentation  |  Google for Developers — https://developers.google.com/search/docs/appearance/ai-features
[7] KPIs & Metrics Explained — AI Visibility Measurement — https://docs.finseo.ai/getting-started/kpis
[8] Characterizing Web Search in the Age of Generative AI — https://aclanthology.org/2026.findings-acl.526.pdf
[9] Interpret Answer Engine Insights | Profound Help Center — https://help.tryprofound.com/articles/6240000968-interpret-answer-engine-insights
[10] Introducing Search Generative AI performance reports in Search Console  |  Google Search Central Blog  |  Google for Developers — https://developers.google.com/search/blog/2026/06/gen-ai-performance-reports
[11] What's new in Google Analytics - Analytics Help — https://support.google.com/analytics/answer/9164320?hl=en
[12] Default channel group - Analytics Help — https://support.google.com/analytics/answer/9756891?hl=en
[13] Campaigns and traffic sources - Analytics Help — https://support.google.com/analytics/answer/11242841?hl=en
[14] [GA4] About modeled key events - Analytics Help — https://support.google.com/analytics/answer/10710245?hl=en&ref_topic=14710097
[15] Publishers and Developers - FAQ | OpenAI Help Center — https://help.openai.com/en/articles/12627856-publishers-and-developers-faq%23%3A~%3Atext%3DChatGPT%20automatically%20includes%20the%20UTM%20parameter%20utm_source%3Dchatgpt.com%20in%20referral%20URLs%2C%20enabling%20clear%20tracking%20and%20analysis%20of%20inbound%20traffic%20from%20ChatGPT%20search%20results?bt_alias=eyJ1c2VySWQiOiAiZTM3ZTE3NjUtMmM5Yi00NjZjLWI0NTQtZWQxYzM0ZjY1NzNhIn0%3D
[16] The AMEC GEO Principles — https://amecorg.com/wp-content/uploads/2026/05/AMEC-GEO-Principles.pdf
[17] NEW Research: AIs are highly inconsistent when recommending brands or products; marketers should take care when tracking AI visibility - SparkToro — https://sparktoro.com/blog/new-research-ais-are-highly-inconsistent-when-recommending-brands-or-products-marketers-should-take-care-when-tracking-ai-visibility
[18] How Much Measurement Does an AI Visibility Claim Need? Replicate Instability and Minimum Detectable Effects in Generative Search Audits | Zenodo — https://zenodo.org/records/22480734
[19] Metrics Overview — https://docs.peec.ai/metrics-overview
[20] SEO Contracts: What to Check Before You Sign — https://thefractionalseo.com/blog/seo-contract
[21] Generative AI performance report (Search) - Search Console Help — https://support.google.com/webmasters/answer/16984139?cachebust=20260909&hl=en
[22] Scopes of traffic-source dimensions - Analytics Help — https://support.google.com/analytics/answer/11080067?hl=en
[23] Decoding Compensation Models & Implementing the Right Model — https://www.aaaa.org/wp-content/uploads/4As-ANA-Decoding-Compensation-Models-and-Implementing-the-Right-Model.pdf?access_pid=118450
[24] Service Agreements for UK Digital Marketing Agencies | Sprintlaw UK — https://sprintlaw.co.uk/articles/what-to-include-in-a-uk-digital-marketing-agency-service-agreement
[25] How to measure AI share of voice using Semrush — https://www.semrush.com/blog/how-to-measure-ai-share-of-voice
[26] Custom channel groups - Analytics Help — https://support.google.com/analytics/answer/13051316?hl=en
[27] 74002 v1.pdf, page 9 @ Preflight ( 74002 v1.indd ) — https://ipa.co.uk/media/4554/agency-remuneration-best-practice-guide.pdf
