GEOChoosing a Contractor

Who to Trust with GEO Promotion: Test-Driving an Agency Before Committing a Large Budget

Oshmanovsky A.O.
A.O. OshmanovskyAgency Director
19 min read
387

What to request from a GEO agency before signing a contract: analysis of an AI response, the case’s source figures, the pilot scope and a full estimate. Tasks for comparing candidates.

Who to Trust with GEO Promotion: Test-Driving an Agency Before Committing a Big Budget

Three agencies have submitted proposals for GEO. One defines success as brand mentions, another sells a package of articles, and the third offers monitoring. You can compare monthly fees right away, but to understand who will change the website and take responsibility for the result, you need to break down the scope of work.

For selection, give the candidates the same input data: the product, geography, customer questions, and one AI-generated answer. In the meeting, ask them to open the sources, explain the problem they found, and propose the first improvement. Then compare pilots that are limited in time and scope.

This article brings together assignments for candidates, questions to ask about case studies, and a framework for comparing estimates. The international example is cited with a link to the authors; the budget calculation and selection scenarios are illustrative.

1. Start with one task you can show on screen

The phrase “we need GEO” gives an agency too much freedom. It can choose convenient questions, show a few mentions, and formally fulfill its promise. Define a situation that affects your customer’s choice.

For example, an equipment manufacturer wants to appear in comparisons of solutions for a small workshop. A regional clinic has found that an AI tool is showing an old address. A corporate portal developer gets mentions, but is described as a studio for small websites. These are different tasks: being present in the selection process, correcting a fact, and changing an inaccurate perception of specialization.

Example brief for a window company

We are promoting glazing for country houses in Moscow and the Moscow region. The priority is projects with large openings. We need to research which companies and solutions appear in answers to questions about this type of glazing and propose a verifiable pilot. We have a technical specialist, project photos, and a developer available for four hours a week. The agency will handle material preparation and coordination.

Add two more constraints. First, what you are not prepared to promise the customer: for example, a fixed price before measurement or manufacturing any structure without a feasibility check. Second, what pilot result the manager needs in order to decide on the next stage. This could be a confirmed problem in the brand description, implemented changes, and a repeat measurement across an agreed sample.

2. Give three candidates the same short assignment

You should not demand a full strategy for free. For selection, a limited review is enough: one product, three buyer questions, and one landing page. The candidate should show their thinking, not complete the project before a contract is signed.

Send everyone the same set of materials: a product link, geography, several real questions for managers, and one example of an AI-generated answer. Do not share your own theory about the cause in advance. Otherwise, you will get three confirmations of your hypothesis.

Assignment for the candidate

In a 20-minute meeting, review one AI-generated answer about our product. Show what can be established from the answer itself and its sources, what requires access to data, and which change you would suggest testing first. Name the person responsible, the work result, and how it should be accepted. If there is not enough data, explain exactly what is missing.

During the meeting, ask them to open the source referenced in the answer. Suppose the AI says the company operates in only one city, while you deliver across the region. The candidate should check where that geography came from: the contacts page, an old listing, a third-party review, or the model’s own assumption.

The right action depends on what is found. If the site contains outdated copy, it can be corrected. If the error is on someone else’s page, you need to clarify the update terms with the owner. If the source is not disclosed, you cannot confidently promise that changing a specific paragraph will solve the problem.

What the candidate does How to evaluate the answer
Shows the source and the exact discrepancy The cause is at least partially confirmed
Separates observation from assumption Understands the limits of the available data
Suggests a specific URL and change Can formulate an executable task
Names the dependency on your team Takes real implementation into account
Promises to fix the answer across all systems after a single publication The basis for such a promise needs to be checked

3. Break down one case study to the raw numbers

In a case study, first check the basis for the calculation: how many questions were tracked, how many answers were received, and whether the sample changed. Then compare the measurement dates with the published website changes.

Ask them to walk you through one project similar to yours in terms of sales type. For a window company, experience with local geography, technical product selection, and a long approval process is useful. Results from an international SaaS project may suggest a research method, but they will not forecast the number of glazing leads.

Start with three questions: what was counted, on what sample, and over what period. Then ask to see one primary fragment: an anonymized log entry, a chart with dates, or a list of changed pages. Access to another company’s CRM is not needed for this.

Training example: why “200% growth” explains very little

In the first measurement, the brand appeared in 2 out of 40 answers; in the second, in 6 out of 40. The number of mentions increased by 200%; the share rose from 5% to 15%, or by 10 percentage points. There may be a useful improvement behind this, but the absolute scale is still small. Without repeated measurements, it is also difficult to assess how stable the result is.

Another case: at the start, 40 general questions were used; by the end, 60 queries containing the company name had been added. The overall share of mentions changed, but this is now a different sample. To assess the trend, the old set of questions and the new expanded set need to be shown separately.

How to check experience with a Russian audience

Take one commercial question with geography, for example: “which windows should I choose for an apartment near a noisy road in Moscow?” Ask the candidate to show the full answer, the date of the check, the available region settings, and the sources. Then compare the answer with what the company actually sells: whether it serves that district, offers the right solution, and supports the specifications with documents.

Training scenario: there is a mention, but no buyer

The agency shows that a window brand appears in an answer to a question about sound insulation. However, the link leads to a general article with no configurations, and the recommended product is not available in the required region. Formally, visibility has increased. For the buyer, the task remains unresolved. The next assignment for the contractor is to find the source of the inaccuracy, verify the facts with a technical specialist, and propose a suitable landing page.

Ask to see how this analysis will appear in the report. It should have separate fields: brand mentioned, link present, description correct, offer suitable for the geography and the task. That way, a positive metric will not hide an incorrect recommendation.

Check what the contractor counts as an inquiry

Training example: three events instead of three customers

A visitor from an AI service opened a form, clicked the submit button, and saw the thank-you page. If analytics counts each action as a separate goal, the report may show three completions. In the CRM, however, there will be one inquiry. Ask the contractor to go through this path and show how they exclude duplicate counting.

For a business, it is useful to separate visits, submitted forms, unique inquiries, and qualified leads. Clarify where the quality check takes place: in analytics, in the CRM, or through a manual reconciliation with a manager. If there is no sales information yet, the report should stop at the last confirmed stage.

Western example: Graphite and Webflow

The Graphite/Webflow case study reports share of voice growth from 33% to 64% and an increase in the share of registrations attributed to LLMs from 4% to 8%. By registration, the authors mean the creation of a free account. Individual experiments are described with a control group taken into account; for example, in the table of contents test, the stated effect differs from a simple change in the initial averages. Source: Graphite (opens in a new tab).

A useful question for the candidate: “Is your percentage a simple before-and-after comparison, or a result adjusted for a control group? Show the calculation.” And one more: “Did you grow in the absolute number of registrations or in the channel’s share?” Both figures can be important, but one must not be substituted for the other. Someone else’s high conversion rate for free registrations is not a benchmark for an expensive B2B deal.

4. Ask them to explain where observation ends and interpretation begins

A GEO report can look convincing even if the data collection method changes every month. Before starting, review one row of the future report: what question was asked, through which interface, with which settings, and what was counted as brand presence.

An API run and an answer in the user interface should be recorded separately. It is important to know whether internet search was used, whether the conversation was new, which language was selected, and whether regional settings were applied. Not all parameters can be fully fixed; the limitations should be documented in advance.

There is also the issue of variability. A SparkToro and Gumshoe study published in January 2026 examined the repeatability of brand and product recommendations. The authors found high variability in the lists and cautioned against directly applying familiar ranking logic to AI-generated answers. Source: SparkToro study (opens in a new tab).

Ask how the candidate accounts for variation between repeated responses. The report should keep individual observations: they show whether the brand appears consistently or whether the average increased because of a few successful runs.

Example of checking a report

The set contains 30 questions with three repeats — 90 planned observations. Ten runs ended in a technical error. The brand appeared in 18 of the 80 responses received. The report should show both 22.5% of the responses received and coverage of 80 out of 90 runs. Technical errors must not be silently counted as the brand being absent or quietly removed from the report.

If some questions do not trigger a generative response, agree on a separate status for that observation. “There was no AI answer,” “there was an answer without the brand,” and “the answer could not be collected” describe different events.

For Russian projects, it is useful to compare your own observations with available Yandex data. The “Site Visibility in Alice AI” tool shows representation among sources, is updated weekly, and contains data for the last three months. Its sample is limited to queries for which the site already ranks high enough in search. Therefore, its percentage cannot be directly equated with the percentage in an agency’s arbitrary set of prompts. Source: Yandex Webmaster (opens in a new tab).

5. Check who will actually make the changes

Suppose the agency has found a problem: the equipment page does not show the range of permitted loads, and an old PDF lists a discontinued model. That is a useful finding. Now find out who will obtain the confirmed specifications, rewrite the page, replace the file, and check the published version.

In the meeting, draw a five-step chain: research → obtain facts → prepare → implement → approve. Put a name or role next to each step. An empty cell means a future delay, not a free add-on service.

Action What to ask before signing the contract
Obtain product specifications Who interviews the expert and records the approved information?
Prepare the material Who writes, edits, and checks the figures?
Make changes in the CMS Is publication included, and who is responsible for layout and images?
Fix a technical issue Is there a developer and an agreed time reserve?
Check the result Who opens the page and checks it against the task?

On the page for Model A, replace the outdated specifications table with data from the approved specification version 3. Add application limitations and a link to the document. Update the related comparison page. Responsible for facts — the client’s engineer; preparation and publication — the agency; acceptance — jointly. Result: a list of URLs, published changes, and a link check.

This work can be accepted regardless of whether a mention appears the next day. The dynamics of responses are then analyzed as a separate result. If the contractor takes on only the preparation of recommendations, this should be reflected in the estimate and timeline.

6. Review the proposal for external publications using one platform

The phrase “we will place the brand on authoritative resources” is too vague to buy. Ask for an example of a platform and an explanation of why it suits your audience, what topic the editorial team will accept, and what publication terms are known.

You need specific answers: whether it will be an interview, a review, an expert comment, or original research; who will confirm the facts; who will approve the material; how the placement cost is recorded; and what happens if the editorial team rejects it.

Western research provides grounds for studying external presence, but not for buying any mention available. In an analysis of 75 thousand brands, Ahrefs found a correlation of 0.664 between web mentions of a brand and visibility in AI Overviews. The authors emphasize separately that correlation does not prove causation. These data cannot be used to promise that “five publications will deliver such-and-such percentage growth.” Source: Ahrefs research (opens in a new tab).

Example of a specific agency response: “In the current sample, two reviews on this platform regularly appear among the sources. We have a comparison topic that your engineer can confirm. First, we will clarify the editorial terms; inclusion of the material in AI answers is not guaranteed.”

Example of a weak response: “We’ll buy a hundred mentions across our network of sites; neural networks like scale.” Here, neither the audience, nor the quality of the pages, nor the connection to the buyer’s choice is clear yet.

It is not worth mechanically transferring a Western list of platforms into a Russian niche. Reddit or English-language YouTube may be relevant for an international IT product. For local manufacturing, first research the sources that actually appear for your customers’ Russian-language questions. When choosing platforms, also consider the ordinary value of the publication: whether people will be able to find, read, and use it even if an AI system does not cite it.

7. Compare the total cost of the same pilot

Two proposals may have the same monthly payments but different implementation scopes. To compare them, set the same boundaries: one product, one geography, an agreed set of questions, several pages, and a defined timeframe.

Ask to split the budget into team work, data collection, development, content preparation, and external expenses. Specify which tasks you will handle yourself. This will show where a lower price is explained by a smaller scope and where by a more convenient setup.

Training calculation. The figures below are fictional and are not market rates of Enterno or other agencies.

Pilot cost item Proposal A Proposal B
Agency work 120 000 ₽ 180 000 ₽
Development billed separately 40 000 ₽ Included within the agreed scope
Editing and publishing content 30 000 ₽ Included
Monitoring tool 10 000 ₽ Included
Total comparable external costs 200 000 ₽ 180 000 ₽

With these inputs, Proposal B is 20,000 ₽ cheaper, even though its base price is higher. But the conclusion will change if you already have an available editor and developer, and the separate costs in Proposal A are not required.

Do not calculate the cost per lead based on an assumed percentage of AI visibility. Between an AI-generated answer and a sale, there is a click-through, a landing page, an inquiry, qualification, and the manager’s work. Until these stages are measured, the estimate describes the cost of the pilot, not a guaranteed acquisition model.

8. Buy a pilot with a clear endpoint

Before signing an annual contract, run a limited project. Agree on the dates for diagnostics, publications, and the final meeting; the resources available from your team; and the materials you will receive. At the final meeting, decide which activities to scale and which observations to continue tracking.

You can allocate six weeks for the first cycle. This timeframe applies to the team’s work. Indexing, source recrawling, and the accumulation of inquiries may take longer. That is why you should separate acceptance of the completed work from assessment of the market impact.

For example, a window manufacturer has three priority scenarios: a noisy street, a child’s room, and panoramic glazing. The pilot can cover one scenario, two landing pages, and several relevant external sources. This makes it easier to understand why each task was chosen than if you tried to overhaul the entire website at once.

Stage What the team does What the client retains
Week 1 Clarifies the product and buyers’ questions, captures the baseline answers, checks analytics A list of queries, saved answers, a source table, and a description of the measurement method
Week 2 Analyzes the reasons for the gaps and agrees on priorities A plan of specific changes with owners and dates
Weeks 3–4 Publishes materials, implements technical changes, and starts the agreed publications Page URLs, content versions, and an implementation log
Week 5 Checks whether the changes are accessible and repeats the measurements Recrawl data where available, new answers, and identified errors
Week 6 Compares observations with the baseline and reviews inquiries A report with limitations, costs, and a recommended next step

If an external platform does not manage to publish the material in time, the report should show the actual status: “approved,” “sent to the editorial team,” or “published.” A submission sent to an editorial team is not the same as an earned mention. The same applies to a developer’s task: a ticket is closed only after the change appears on the live website.

9. Review one uncomfortable number at the final meeting

Let’s imagine a sample pilot result: the share of answers mentioning the brand increased from 10 to 25%, traffic grew, but the number of qualified inquiries stayed the same. This is not yet a reason either to fire the agency or to double the budget. First, you need to understand what happened between the answer and the lead.

Ask the team to walk through one path end to end: user question → answer → referenced page → action on the website → inquiry record. Each transition may reveal its own problem.

Observation What to check next
The brand is mentioned, but there is no link to the website Track mentions and citations separately; do not promise a direct click-through from every answer
The link leads to a general article, although the question is about making a purchase Check whether the page includes the right product, region, and next step
There are visits, but the form is barely opened Review the relevance of the page, the mobile journey, the terms, and form availability
Goals are being tracked, but the manager does not see any inquiries Check the analytics events, form submission, and receipt of the inquiry in the CRM
There are inquiries, but the product or geography is not a fit Review the wording of the questions and the promises made in the materials

Do not expect a small sample to deliver the accuracy of a large advertising channel. Two inquiries instead of one may simply be random fluctuation. To evaluate sales, you need both sufficient volume and time for the deal to close. But even with low traffic, a contractor can show what data has been collected, where the journey breaks down, and which change is worth testing.

Review discrepancies between systems separately. Visits with a recognizable AI referrer are only the visible part of the traffic. A user may later type the address manually, switch from another device, or arrive via search. At the same time, you cannot label all direct traffic as “hidden GEO.” If the agency uses modeling or a “how did you hear about us?” survey, that data should be shown as a separate line item.

A good proposal for the next step sounds specific: “Questions about sound insulation are now bringing up links to a review. Visitors cannot find equipment options or measurement terms in it. We will add a comparison of two suitable solutions, track clicks through to the inquiry form, and repeat the observation.” An explanation like this lets you assess the logic of the decision and the scope of the new work.

10. Choose a team by scores — and by reasons to say no

After each meeting, fill in the same evaluation table and record the reasons for your scores. Otherwise, a detailed analysis and a confident presentation may leave the same impression. Below is a working scale; you can adjust the weights to fit your project.

Criterion Maximum What deserves a high score
Understanding of the product and demand 20 The team distinguishes your buyers from the broader audience interested in the topic and explains its choice of questions
Measurement quality 25 There are source responses, comparable conditions, repeat measurements, and transparent denominators
Ability to implement 25 The team names the people responsible, shows real tasks, and makes CMS limitations and timelines clear
Work with sources and facts 15 Platforms are selected based on the task, claims are verified, and independent publications are not faked
Budget transparency and handover 15 Additional costs, access rights, the contents of the export, and the project closure process are visible

For each point, use a simple rule: a promise without an example gets a low score; an example without an explanation gets a medium score; a verifiable example with clear logic gets a high score. Write the reason next to it: “showed the measurement log,” “did not name the editor,” “development is excluded from the price.” The total score helps compare proposals, but it does not replace these comments.

There are reasons to decline regardless of the score. For example, a candidate proposes fake reviews, promises a permanent first place in answers, hides source data behind the phrase “proprietary methodology,” or presents someone else’s result as their own. A strong presentation score does not compensate for conditions like these.

The Western Surface Labs guide for GEO clients (opens in a new tab) makes a separate distinction between an agency, a monitoring tool, and an internal team. For procurement, this is a useful fork in the road: a measurement service may suit a company with its own editors and developers, while another business may need hands-on contractors. Surface sells its own platform, so its comparison should be read with that commercial interest in mind.

Apply this fork in the road to your own situation. If your SEO team already produces high-quality materials and can quickly update the site, you may need a specialist in measurement methodology and source analysis. If materials sit in approvals for months, another monitoring tool will not solve the problem: the pilot will need someone responsible for publication.

11. Agree from the start on what will remain after the pilot

Even a good project can come to an end: the budget, season, or product priorities may change. Make sure that in two months, another employee can continue the work without having to reconstruct the history from email threads.

Before the start, list the materials to be handed over: queries and measurement conditions, source responses, source URLs, tasks, published texts, a list of site changes, and the final export. If the tool does not allow all fields to be exported, find this out in advance and agree on an available format. Screenshots of charts are usually not enough for repeat analysis.

Grant access to analytics, the CMS, and tools personally, and only to the extent required for the tasks. Viewing reports does not always require administrator rights. When the project is complete, it should be clear which accesses must be revoked and which integrations will continue to work. Service tokens and passwords should not be left inside a shared presentation.

Clarify who stores the editorial source files, pays for monitoring renewals, and maintains the published materials. A paid publication may have a limited placement period, and a tool may stop showing history after the subscription is canceled. These details affect the cost of continuing the work.

A short brief you can start with tomorrow

You do not need to compile a long tender document. Send candidates the same description and ask them to respond to the point:

We are promoting [product] for [buyer] in [geography]. The priority selection scenario is [client task]. We currently have [website, analytics, team, and materials]. Using one question, show how you will capture the baseline situation, find the gap, and choose the first change. Propose a limited pilot: scope of work, team members, full cost, data to be handed over, and criteria for deciding whether to continue. Separately, specify which results you can accept as completed work and which depend on external systems.

Next, hold three meetings, fill in the comparison table, and check one case study from each finalist. The choice will become easier: you will have specific differences in methodology, team, and cost.

Before signing the contract, you should have an agreed list of questions, a measurement method, named team members, a budget with additional costs, and an acceptance procedure. If any of these points remains unclear, return to it in the negotiations: after the start, a disagreement will cost more.

At Enterno, you can start by discussing your product and one priority neural search scenario. Tell us about your GEO promotion task — and we will discuss what data and work are needed for the first stage. And if you would first like to understand the technology itself, read our practical plan for promotion in neural networks.

Oshmanovsky A.O.

Agency Director

Business development strategy and the creation of effective digital solutions.

Share: