GEO consultant: 47 offers read, what to demand

"GEO consultant" is a job title that did not exist two years ago. It sells today in Paris, Lyon and New York, and the buyer signing for it has no idea what is in the box.

So we read the offer pages of 47 vendors, on 2026-08-28: 18 French agencies, 12 French independents, 9 tool vendors, 6 international agencies, one freelance marketplace, and vydera.com, because exempting ourselves from our own grid would have been convenient.

The result fits in one sentence. The market uses eleven different words for a single discipline, 29 of 47 vendors promise the same artefact, 7 publish the price of what they sell, and none commits to how long the work runs.

First surprise: "agence geo" is not a surveyor's query

The brief for this article carried a warning: the French query "agence geo" was probably owned by geomatics and land surveyors, so check the SERP before targeting it. The SERP says otherwise.

Measured 2026-08-28, Google France, French language, desktop: 19 organic results on "agence geo", zero of them geomatics. Fifteen are GEO sales or content pages in the Generative Engine Optimization sense. The other four are directories, agency rankings and a video, all marketing. Same on "consultant geo": 19 organic results, zero geomatics.

The rest of the page confirms it. An AI Overview holds position 1 on "agence geo". People Also Ask includes "What does a GEO audit cost?" and "Which are the best GEO agencies?". Related searches read "Agence GEO Paris", "Agence GEO Lyon", "Meilleure agence GEO". The marketing sense of GEO has taken the whole French SERP on both queries.

Two caveats, because a measurement that hides its holes is not a measurement. The content of that AI Overview was not captured: the API flags it in position 1 but returns it asynchronously, with no text and no references. And the search volume for "agence geo" stays unknown: the tool returns an empty value for it, as it does for "prestation geo", so no volume figure will be written here. Only "consultant geo" is measured, at 210 searches a month, difficulty 0.

How the 47 pages were read

The method matters as much as the result, because it is reproducible.

• 26 vendors come from the two SERPs. The page measured is the one Google ranks, not the one we would have picked for them.
• 21 were added by hand: the French agencies from our E-E-A-T audit panel, six international agencies, nine tool vendors. For those, the offer page is discovered from the homepage, never guessed from a plausible URL.
• Directories and rankings are excluded, however well they rank: they talk about the market, they do not sell the service.
• The unit of count is the vendor, never the item. Counting bullets on a page measures its layout, not the clarity of its offer.
• Everything is read on served HTML, with no browser. Anything injected by JavaScript is invisible, and that is the main limit of the exercise.

Every detection keeps 120 characters of raw context, so the classification can be contested with the evidence in hand. That is what made the price re-reading below possible, and it changed the number.

Eleven words for one job

Eleven distinct terms name the discipline across these 47 pages: GEO (42 vendors), AI visibility (38), AEO (16), AI SEO (14), AI search or generative search (9), French "référencement IA" (7), LLMO (7), GSO (3), SEO ChatGPT (2), GAIO (1), SAIO (1).

The average is 2.98 terms per page. One vendor uses none, six use one, thirteen use two, nine use three, and eighteen out of 47 line up four or more. Passion Référencement, Promoovoir and Semrush stack six each.

A twelfth acronym was dropped from the count after a hand review, and it is the most instructive case in the whole survey. AIO does not have the same referent from one page to the next: at Otterly.AI ("Track ChatGPT, Perplexity & Google AIO") and Terakeet ("monthly AIO mentions", glossed on the spot as "brand mentions in AI Overviews") it means AI Overviews, an engine; at Semrush ("Enterprise AIO") it means the discipline. One acronym, two referents, across eight vendors. Counting it would have added apples to oranges.

This lexical inflation is not a vocabulary detail. It is why a marketing director comparing three proposals is comparing three different words without knowing whether they cover the same work. We laid out the definitions in our piece on AI SEO; the market has not settled anything.

One vendor in the panel, Primelis, carries no term of the discipline on the page read. It is also the only one whose GEO offer page was never discovered from the homepage: the reading covers its homepage alone, and should be read as such.

What the market says it delivers

Each vendor gets one rung, the highest it reaches across its content area. T3: it names an artefact verifiable on delivery, a file, a format, an object. T2: it quantifies a volume or a deadline without naming an artefact. T1: it names activity categories only, audit, strategy, support. T0: it promises an outcome and nothing else.

The count: 36 vendors at T3, 2 at T2, 9 at T1, none at T0. A lazy headline would stop there and say three quarters of the market is specific. It would be wrong, because T3 is a very low bar: the artefact named is nearly always the same one.

29 of 47 vendors cite structured data or schema.org. That is rank one, and it is an artefact SEO has been naming for a decade. Then come the dashboard (16), the entity record or knowledge graph (12), then a plateau at eight vendors: llms.txt, a delivered report, an export or API, sitemap.xml or robots.txt. At the bottom, the content brief (3) and the redirect plan or site structure (2).

The median is 2 named artefacts per vendor. Put plainly: half the market promises structured data, possibly a dashboard, and nothing else you can check.

Two honest caveats on this ranking. The T1 / T3 boundary is arguable when the artefact shows up inside an explanation of how engines work: Origine writes that "AI values the implementation of FAQ schemas", which was read as teaching and leaves the vendor at T1, though a reader may read it the other way. And the page read is not always a vendor's best page: at Terakeet, homepage discovery picked up two case studies rather than a service page, so the absence of artefacts holds for those pages, not for the whole offer.

An item-level extraction was attempted too: 128 items returned under section headings that announce a service, across 27 vendors, none of which carries a countable quantity. That corpus is not a denominator and serves only as quotable material. Yumens and NOIISE write their offer in full paragraphs and return zero items without being any vaguer for it. An earlier, wider pass returned 1,895 "deliverables" across 47 vendors, team photos and CV lines included: that number measured page layout, and it was thrown away.

The promised measurement, the one point of unanimity

On one point the market speaks with a single voice. 46 of 47 vendors name at least one tracked engine, 40 name at least one metric, 43 state a frequency, and 38 name all three. The median is 7 engines cited per vendor.

Engines, by citation count: ChatGPT (46), Gemini (42), Perplexity (41), Claude (33), AI Overviews (31), Bing (25), Copilot (25), AI Mode (16), Mistral or Le Chat (10), SGE (9), Grok (6), DeepSeek (6), Meta AI (1). Metrics: position or rank (29), share of voice (25), brand mentions (21), sentiment (16), cited sources (13), citation rate (12), visibility score (11), AI referral traffic (8). Frequencies: daily (29), monthly (24), real time (23), weekly (13), quarterly (4).

Read that consensus for what it is. Promising measurement costs almost nothing: these lists copy from page to page, and the promise is identical at the agency billing 700 € a day and at the tool billing $29 a month. What costs is handing over the raw data behind it: an export or an API is mentioned by only 8 vendors out of 47, though it is the base brick of any SEO and GEO dashboard.

Price: 15 amounts detected, 7 real prices

The automated rule is simple: a currency amount in served HTML counts as a published price. It finds 15 vendors out of 47, against 6 who explicitly say "on request" and 26 who say neither.

Except the file keeps the context around every amount. Re-reading all fifteen vendors one by one, on 2026-08-28, only seven are the price of the service sold. The other eight are a fundraising round, client case-study budgets, a form dropdown, a fictional company's rate card inside a product demo, the price of a WordPress site in a blog post, or a pair of running shoes in a mock ad.

Both counts are published because both are defensible: 15 under the survey's own definition, 7 under a stricter one, mine, contestable with the evidence in hand. Keep the second: 7 out of 47. Three of them are French, out of thirty French agencies and independents read.

And there is no market median. The amounts mix one-off audits, monthly retainers, software licences and dollar entry tickets: they do not average. The most complete rate card in the panel is SLASHR's, which publishes all three at once: a GEO audit from 1 500 to 5 000 € excluding tax, monthly support from 1 200 €, and a day rate of 600 to 700 €.

One last warning, and it matters: 26 "undetermined" vendors are not 26 vendors hiding their price. The survey reads served HTML only. A rate card injected by JavaScript is invisible to a plain fetch, exactly as it is to most AI engine crawlers.

Commitment: nobody commits to a duration

25 of 47 vendors let you read an engagement format, 22 do not. The formats found: monthly retainer (16), package (8), trial or entry tier (5), sprint or scoped mission (3), day-rate staffing (1), one-off audit (1).

The number that matters sits elsewhere. Duration detection finds only four mentions across 47 pages, and all four say the opposite of a duration. Stride-up writes "Durée indéterminée, sans engagement", open-ended and with no commitment. Yann Dubuisson writes "Prestation mensuelle et sans engagement sur la durée", a monthly service with no term. The other two, Agence SEO GEO and Conversion Boosters, turn "no commitment" into a selling point.

A fifth mention of duration turns up outside that detection, inside a price context: Paul Vengeons publishes "Temps moyen de prestation 3 à 6 mois", an average time on the job of three to six months. It is an observed average, not a commitment, and it is the most precise thing the panel produces on the question.

In other words: across 47 pages selling AI visibility work, exactly one publishes an order of magnitude for duration, and none commits to it. A client who signs buys a monthly fee with no horizon.

What to demand, in this order

1. The word, and why that one. Ask which of the eleven terms your vendor uses, and what it covers for them. If the answer is "they are all the same", you have the market average: 2.98 terms per page.
2. The list of artefacts, named. Not "audit", not "support": a file, a format, an object that lands in an inbox. The market median is two. Below three, you are buying an activity category.
3. Whatever comes beyond schema.org. 29 of 47 cite structured data: that is the floor, not the differentiator. llms.txt, 8 of 47. A content brief, 3. A redirect plan, 2.
4. The engines, named, with the measurement frequency. The panel median is seven engines cited. "All the AIs" means nothing and costs nothing to write.
5. Access to the raw data. Export or API: 8 of 47 mention it. A dashboard with no export is a wall, not a deliverable.
6. The price, in writing, before the call. Seven vendors out of 47 publish it. You are entitled to ask for it, including from a vendor who does not display it.
7. The duration, and what happens at the end. No vendor in the panel commits to one. Which makes it the fastest question to separate two proposals.

These seven points overlap the grid we published for choosing an AI-era SEO agency, with one advantage: each rests on a number you can verify yourself on the market's public pages.

What this measurement does not say

It covers what 47 vendors write on a sales page, not what they deliver. The real quality of the artefacts is not measured, and a T1 vendor may well deliver more than a talkative T3 one.

The panel is not a random sample either: built on two French SERPs plus a manual addition, it over-represents vendors who deliberately position on the word GEO and under-represents those selling the same work without ever using it. Finally, a vendor citing a term in order to dismiss it is counted as using it, except for the four rare terms reviewed by hand.

Where Vydera lands in its own panel

vydera.com is one of the 47, measured on the same grid, and the result is not flattering everywhere.

On the credit side: rung T3, six artefacts named across the pages read, including llms.txt, structured data, a dashboard, a report and the entity record. Three terms used, below the panel average. And the complete measurement promise, engine, metric and frequency, like 38 other vendors.

On the debit side: no price published on the pages read, which puts Vydera among the 40 vendors with no readable price under the strict definition used here. A readable engagement format, the monthly retainer, but no committed duration, like the other 46. On the two hardest demands in the list above, the sixth and the seventh, Vydera behaves exactly like the market. That is a debt, it is written down here, and this article works as much as a summons as a reading grid. The page carrying the offer is our agentic GEO agency.

Running the measurement again

Nothing here needs a paid tool, apart from the two SERPs, billed at $0.0035 each. The rest is a script that downloads a homepage, finds the links whose anchor announces an offer, downloads those pages, and runs regular expressions over the content area. Thirty lines of vocabulary to define, one hour of hand review for the edge cases.

That hour is the part nobody skips: re-reading the contexts halved the number of published prices. A figure produced by a regular expression and shipped without review is a wrong figure that looks right.