Share of answer is the operating metric for this practice.

How we count it
Request an audit

Insight · 13 min

How to choose a GEO agency without buying a screenshot

To choose a GEO or AEO agency, ask for the baseline before you sign the retainer: a fixed set of buyer queries, a named panel of engines, several runs of each query, and mentions, citations, links, and accuracy reported as separate numbers. An agency that can show you that is measuring something. One that leads with a single visibility score, a guarantee, or a plan to seed forums is selling a feeling. As of November 2026 there is no industry standard for this work, its methods, or its price, so the tests below are the only protection a buyer has.

AIEOLABS sells this work. Apply every test in this guide to us too.

In short

  • Most of GEO is SEO. Google says so for its own surfaces. What is new is passage extraction, citation selection across sub-queries, off-site corroboration, and answers that change from run to run.
  • A credible baseline names the queries, the engines, the run count, and the date, and scores mention, citation, link, and accuracy apart.
  • Walk away from guaranteed citations or rankings in AI answers, a single score with no run count, “rankings” on prompts no buyer types, Reddit seeding, and paid links without rel="sponsored" or rel="nofollow".
  • Pricing has three common shapes: a fixed-scope diagnostic, a monthly retainer, and tool subscriptions. Published ranges are vendor estimates, not survey data.
  • You can run a small baseline yourself in a few days. Hire when the work you need is content, technical fixes, and earned coverage at a pace your team cannot sustain.

Why are you being pitched GEO now?

You are being pitched because someone senior saw a competitor named in ChatGPT, and your own brand was missing or described wrongly. That moment is real and worth acting on. It is also one sample from a system that gives different answers on different runs, so it is a reason to measure, not yet a reason to buy.

The trigger tends to arrive with a contradiction: the company ranks well on Google, and the assistant still names someone else. Both can be true at once. A ranking rewards a page. A generated answer rewards a claim that the engine found on the sub-queries it ran, often on sites you do not own. The first question for any agency is whether it can tell those two problems apart for your queries.

What is actually different from SEO work?

Less than most pitches say, and more than nothing. Google’s guide to generative AI features in Search says that “from Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.” AI Overviews and AI Mode draw on Google’s Search index and its ranking and spam systems. Crawl, indexing, canonicals, and content quality carry over to every engine.

What changes is the selection step and the scoreboard:

  • Extraction. The answer lifts a passage, not a page. A page can rank and still contribute nothing.
  • Fan-out. Engines run related sub-queries and cite sources from those. Ranking for the head term does not settle who gets cited. See query fan-out.
  • Corroboration. What other sites say about you carries more of the answer than it carried of the ranking. See why AI cites third parties.
  • Variance. The same query returns different brands and sources on different runs.
  • Measurement. Google and Bing report AI impressions and Copilot citations. ChatGPT, Perplexity, Claude, and Gemini give site owners no report at all.

The full argument is in SEO, AEO, and GEO are one system, and the terms are defined in SEO vs AEO vs GEO. For a buyer the practical consequence is simple. An agency that cannot talk about crawl and indexing is not ready for GEO. An agency that talks only about crawl and indexing is doing SEO under a new name, which is fine if that is what you need and priced like it.

What should a credible baseline contain?

A baseline is the measurement taken before any work starts, so later movement can be compared to something. It should contain six things, and you should be able to read each one in the deliverable.

ComponentWhat it must containAsk to see
Query universeA frozen list of buyer queries, reviewed by someone who sells, tagged by intent (commercial, definitional, procedural, comparative). Branded queries labeled and reported apartThe list itself, with the date it was frozen
Engine panelThe specific products tested (for example ChatGPT search, Google AI Overviews, AI Mode, Perplexity, Gemini, Copilot, Claude), and why thoseEvidence the panel matches what your buyers use: referral data, customer interviews
Run protocolRuns per query, number of days, location, language, device, logged-in or logged-out, consumer interface or APIThe run count per query per engine, in writing
Four scores, apartMention, citation, link, and accuracy, each as the share of runs in which it occurredA table with four columns, not one number
Accuracy referenceA dated record of your current prices, plans, and claims, used to grade what the answers sayWhich answers were wrong, and which source each error came from
Raw logOne row per query, engine, and run: answer text, sources, date, screenshotA sample of rows you can re-check yourself

The method behind each line is set out in how to measure AI visibility, and the headline metric that comes out of it is share of answer.

Repeat runs are the item most often skipped. A 2026 paper on measuring AI search makes the point in its title, “Don’t Measure Once,” and notes that “answers can vary across runs, prompts, and time, making one-off observations unreliable.” If two months differ by less than the run-to-run spread, nothing changed. A baseline without a run count cannot tell you that.

The baseline should also use the first-party data that exists. Google’s Search Console has a Generative AI performance report that shows impressions in AI Overviews and AI Mode, without queries or clicks. Bing Webmaster Tools reports citations in Copilot and Bing’s AI summaries. An agency that ignores both and shows you only its own dashboard is choosing not to check its numbers.

What are the red flags?

The red flags are claims no one can deliver and tactics that break a platform’s written rules. Each one below is checkable against a primary source.

Red flagWhy it is a problemWhat the source says
Guaranteed citations, rankings, or “placement” in ChatGPT or AI OverviewsNo outside party controls what a model cites. The answer is sampled and changes between runsGoogle: “No one can guarantee a #1 ranking on Google,” and for AI search, “No third-party tool has access to our internal ranking or AI systems”
A single visibility score with no run countOne number blends mention, citation, link, and accuracy, and hides whether a change is larger than the noiseResearch on AI search measurement recommends treating visibility as a distribution, not a single observation
“AI rankings” on prompts no buyer askedPrompts that contain your brand name, or that were written to produce a win, measure recall of the prompt, not your visibility in the categoryAsk where each prompt came from. Synthetic prompts are a known limit of every tracker
Reddit seeding or astroturfingPosting as fake customers, or posting without disclosing the connection, breaks Reddit’s rules and can break consumer-protection lawReddit’s rules: “Participate authentically,” no “content manipulation,” and “do not intentionally mislead others.” Google says “seeking inauthentic ‘mentions’ across the web isn’t as helpful as it might seem.” The FTC’s 2024 rule bars buying fake reviews and restricts undisclosed reviews by company insiders
Paid links or paid posts without link attributesGoogle treats buying links for ranking as link spam unless the link is qualifiedGoogle’s spam policy lists “Exchanging money for links, or posts that contain links,” and says paid links are not a violation “as long as they are qualified” with rel="sponsored" or rel="nofollow"
A page for every prompt or fan-out variantMass-produced pages written to catch sub-queries are a known spam patternGoogle says creating content for every fan-out variation “primarily to manipulate rankings or generative AI responses” violates its scaled content abuse policy
llms.txt, “AI files,” or schema as the headline deliverableUseful hygiene at best. Not a citation lever on any published evidenceGoogle: Search does not use llms.txt, and there is “no special schema.org markup you need to add.” See llms.txt
Secrecy about the workYou cannot evaluate changes you are not shown, or undo ones that cause harmGoogle: “Be careful if a company is secretive or won’t clearly explain what they intend to do”

One flag deserves a sentence more. Reddit and forum seeding is the tactic most often sold under GEO, because forums are visible in AI answers. It is also the one most likely to embarrass you, since moderators, users, and platforms remove undisclosed promotion, and the posts stay attached to your brand name. Earned mentions last for the same reason seeded ones get removed: someone independent chose to write them.

What questions should you ask before signing?

Ask questions whose answers you can check. The good answers below are specific and falsifiable. The weak ones are not wrong in every case, but they leave you with nothing to hold the agency to.

QuestionGood answerWeak answer
How many times do you run each query, on which engines, and from where?A number per query per engine, over more than one day, with location, language, and account state named“We track thousands of prompts daily” with no per-query run count
Can I see the query list before you start?Yes, drafted with our sales team, frozen and versioned, branded queries kept apart“Our tool generates the prompts”
What do you report?Mention, citation, link, and accuracy as separate shares, plus Search Console and Bing data alongsideOne AI visibility score, or a share of voice with no definition
How will you know the work caused a change?A baseline, an observation window, a log of every concurrent change, and movement read against run-to-run spread“You will see the screenshots”
What happens when an answer gets our brand wrong?Trace the claim to its source, fix what we control, and request corrections where we do not. See when AI gets your brand wrong“We will optimize your content so the AI learns the truth”
How do you earn third-party mentions?Original data, expert commentary, review programs, and PR, with disclosure where there is a connection“We have a network of accounts” or “we place posts”
What will you change on our site, and will we see every change?A written list of changes, owned by named people, reversible“Our process is proprietary”
What technical checks come first?Crawler access by token, CDN rules, rendering, canonicals, indexing. See robots.txt for AI crawlersNo mention of crawl or index
How do you connect this to pipeline?AI referral sources in analytics, self-reported attribution, branded search trends, with limits stated. See AI search attribution“AI traffic converts several times better” with no source
What can you not control?Model updates, sampling, what third parties write, which engines your buyers use“Nothing, our system handles it”

Last, ask who does the work. A named lead you meet before signing is a better predictor of what you get than the deck.

How is GEO work usually priced?

There is no industry standard for GEO pricing, and no independent survey of it that we could verify as of November 2026. What exists are vendor price pages and estimates published by agencies that sell the work. Work is commonly sold in three shapes.

ShapeWhat you getHow it is usually pricedWhat to check
Diagnostic or auditA baseline, a technical review, and a prioritized plan, over a fixed periodA fixed fee for a fixed scopeThat the baseline in the table above is a deliverable, not a sales step
RetainerOngoing measurement plus content, technical fixes, and outreachA monthly fee, often with a minimum termWhat is in scope each month, and what the report contains
ToolsAutomated prompt tracking across a set of enginesA subscription, usually by prompts, engines, and frequencyRun count per prompt, interface (consumer product or API), and location settings

Tool pricing is public for some vendors. As of November 2026, Peec AI lists plans from $95 to $495 a month, priced by tracked prompts and models, with answers counted as prompts × models × frequency, and custom enterprise pricing above that. Profound lists a free trial and custom enterprise pricing without published monthly figures. The pricing model tells you what drives cost: more prompts, more engines, and more frequent runs.

Agency pricing is mostly quoted, not published. Nutshell, a site run by the agency WebFX, gives an estimate of $1,500 to $50,000 a month for GEO agency services, by company size, based on its review of provider pricing. Treat that as one seller’s range. The spread exists because the scope does: a monitoring retainer and a program that produces original research and earned coverage are different products.

AIEOLABS does not publish prices in this guide. The engagement models describe the three shapes we sell and what each includes.

Should you do it yourself or hire an agency?

Do the baseline yourself if you can spare a few days. Hire when the work that follows needs more hours or skills than you have. The measurement is the cheap part. The fixes are the expensive part.

Your situationLean toward
You need to answer “are we visible?” for leadership, onceDIY. A 50 to 100 query set, three to five runs per query on three engines, logged in a spreadsheet
You have an SEO team that already ships technical and content workDIY measurement plus a tool, or a one-off diagnostic to check your method
The problem is crawl, rendering, or indexingYour SEO team or developers first. That work is SEO
Answers describe your product wrongly, and you cannot trace where the errors come fromA diagnostic, from someone who will show you the sources
The gap is third-party coverage: reviews, comparisons, analyst and press mentionsAn agency or PR partner with a disclosed, earned approach
You need a quarterly report that leadership trusts, across several engines, without building itA retainer or a tool, judged by the baseline tests above

Whichever way you go, the query set should be yours. Write it with sales, keep it, and give it to any vendor you hire. If you change vendors, the history stays comparable.

A worked example: the screenshot from the CEO

This is a hypothetical, written to show the sequence. It is not a client story.

A B2B software company’s CEO forwards a screenshot. ChatGPT, asked for the best tools in the category, named three competitors and not the company. A GEO agency has already sent a proposal for a monthly retainer, with a guarantee of “top placement in ChatGPT within 90 days.”

The marketing lead does three things before replying.

  1. Reproduces the observation. She writes 40 queries with the sales team, including the CEO’s exact prompt, and runs each five times in ChatGPT, Perplexity, and Google AI Mode, logged out, over two days. The brand appears in a minority of runs on the CEO’s prompt, and in most runs on procedural queries where the company’s documentation is cited. The screenshot was one of the bad runs on the weakest query type.
  2. Splits the scores. Mentions are low on comparison queries. Citations of the company’s own pages are reasonable on how-to queries. Two answers state a price the company retired last year, sourced to an old review page.
  3. Reads the proposal against the tests. The guarantee is a red flag on its own. The proposal shows one visibility score and does not say how many runs it uses.

The reply to the CEO is a one-page note: where the brand appears, where it does not, the wrong price and where it comes from, and a plan focused on comparison content and third-party coverage. The agency decision becomes a choice between vendors that can show a baseline like hers. That is a better position to buy from than a screenshot.

How do you apply these tests to AIEOLABS?

The same way. Ask us for the run count, the engine panel, and the query list. Ask to see the four scores apart and a sample of the raw log. Ask what we cannot control; the answer is model updates, sampling, and what other people write about you. We do not guarantee citations or rankings, and you should be wary of anyone who does, including us if we ever start.

Our engagement page commits to a shared query universe across SEO, AEO, and GEO, absences listed with a reason, inaccurate answers treated as incidents, crawler policy as a written decision, and meeting the people who do the work. Those are our claims. Hold us to them the way you would hold anyone else. The measurement we use is described on the platform page. If you want to put the questions to us directly, tell us the domain.

Common mistakes

  • Buying from the screenshot. One answer is one sample. Reproduce it across runs and engines before you decide anything.
  • Accepting a blended score. Ask for mention, citation, link, and accuracy apart, or you will not know which one moved.
  • Letting the vendor own the query list. Write it with sales and keep it. Otherwise the vendor can change the test and call it progress.
  • Skipping the technical check. A page the AI crawlers cannot fetch or read cannot be cited, whatever the content plan says.
  • Paying for placements. Paid links and posts need rel="sponsored" or rel="nofollow", and undisclosed forum posts break platform rules. Earned coverage is slower and keeps.
  • Treating tool output as ground truth. Trackers sample, often through APIs, on prompts someone wrote. Spot-check against manual runs.
  • Ignoring the first-party data. Search Console and Bing Webmaster Tools are free and are the only reports the engines themselves provide.

Sources

Bring us the domain.

We will tell you where you are absent, which page is in the way, and what has to be true before a citation is possible.

Request an audit