How B2B Companies Should Choose a GEO/AEO Agency

Knowing how to choose a GEO/AEO agency starts with how the agency measures results. A strong agency records where a company stands in AI answers before any work starts, then reports every change against that baseline. Generative Engine Optimization (GEO) and Answer Engine Optimization (AEO) are closely related names for that work, which improves how ChatGPT, Claude, Perplexity, and Google's AI features mention, cite, and recommend a company.

AI answers now sit early in software research. In G2's March 2026 survey of software purchasers, 51% start research with an AI chatbot more often than with Google.

For a B2B SaaS, fintech, or AI company, the risk is paying for work whose impact cannot be checked. This guide shows what a GEO agency does, when to hire one or wait, how to judge a proposal on five checks, and which questions to ask before signing.

Key takeaways

• Judge the measurement first, because answers shift between runs and a dated baseline with repeated readings of fixed questions makes a change inspectable.
• Track questions that do not name the company, taken from sales calls, with mentions, citations, and recommendations counted apart.
• Ask for one dated case with raw answers, since a rank position or a single screenshot proves little.
• Write the baseline, a 90-day review, and data ownership into the contract, along with what any guarantee covers.
• Fix broken SEO basics and slow approvals first, keep GEO in-house when the team has the people, and wait while prospects rarely use AI answers.

What does a GEO or AEO agency do, and why does the choice matter?

Beyond the best-known AI assistants and Google's AI search features, a GEO agency's work also reaches Gemini, Copilot, and any other AI platform prospects use, and its results take three forms in an answer:

• Mention: The company's name in the answer
• Citation: A source link in the answer, with links to the company's own pages counted apart
• Recommendation: The answer telling the reader to pick the company

Together, these make up a company's AI visibility: the share of answers to a fixed set of prospect questions that mention, cite, or recommend it.

The name "GEO agency" comes from a 2023 research paper on Generative Engine Optimization (GEO). Answer Engine Optimization (AEO) began as work on featured snippets and voice answers and now covers AI answers too, so many firms sell the same service as an AEO agency or a GEO/AEO agency. AI SEO and large language model optimization (LLMO) are newer labels for the same overlapping work, as our comparison of GEO, AEO, and SEO explains.

GEO differs from SEO mainly in what it measures: SEO tracks rankings and clicks on search results pages, while GEO tracks mentions, citations, and recommendations inside AI answers. Much of the work itself is traditional SEO and content marketing. What a GEO scope adds is tracking answers to prospects' questions, plus work on the third-party pages and wrong facts those answers draw on.

Being named in those answers matters, as 85% of the 1,076 respondents to G2's survey say they think more highly of a vendor an AI answer includes. Yet TrustRadius found that 94% of technology purchasers who used AI still fact-checked it, so the answer also has to get the facts right.

Three dated changes from the past six months make a proposal easier to check:

• May 15, 2026: Google published a guide to its generative AI features that mythbusts common AEO and GEO tactics.
• June 5, 2026: Google added guidance on third-party SEO tools and services, and its revised hiring guide asks whether AEO and GEO advice matches Google's own.
• August 31, 2026: Search Console's generative AI performance reports reached all websites, a free view of a website's pages in AI Overviews and AI Mode.

How does GEO work change what AI answers say about a company?

In our reading of the documentation and studies, three things shape an AI answer's response to GEO work:

• Access to the company's pages
• The sources an engine reads
• The companies those sources mention

A fourth, the variation in answers between runs, is mostly outside any agency's control and decides how its work can be measured.

1. Access: The engine has to reach and index the page

Crawlability comes first because Google's AI features show pages that are indexed and eligible for a snippet. Google's guide adds a Search Console setting, the Search generative AI control, which includes a website by default.

AI crawlers matter for the other engines, since OpenAI says websites that block OAI-SearchBot, its search crawler, are not shown in ChatGPT search answers, apart from navigational links, which puts technical SEO in any GEO scope, starting with a crawl check in the first audit.

2. Sources: Each engine reads a different mix of pages

Company websites, competitors' included, drew 47% of ChatGPT's citations and 69% of Gemini's in the median category of Profound's July 2026 report.

Earned media, Profound's bucket for third-party editorial coverage, institutional sources, and press wires, made up 30% of ChatGPT's citations. For SaaS and software companies, the median across all engines was 11.4%, and with mixes that different, a fixed package cannot fit every company.

3. Selection: Sources shape who gets named and what answers say

Besides their sources, models also name companies they already know without citing a supporting page. In a May 2025 study of 75,000 brands, Ahrefs found that web mentions correlated with visibility in AI Overviews far more strongly than backlinks did (0.664 against 0.218). The study shows an association but does not prove causation.

More than half of brands, 54% in Profound's FactCheck data, had at least one wrong AI claim citing their own content, and pricing and billing was the most frequent topic. Most wrong claims cited competitors' pages or earned media, though, so the cited page needs checking before the fix is chosen. An outdated company page gets corrected, and a wrong outside page gets a correction request.

A fintech that cuts a transfer fee can update its own help page the same day. A comparison article or a review that still shows the old figure can keep ChatGPT quoting it. The scope should list each of those external pages and name who will ask its publisher to fix it.

4. Variance: The same question gets different answers across runs

Engines write each answer fresh and change with model releases. A university audit by Uberti-Bona Marin and colleagues of 117 queries about physical products, asked from the Netherlands in September 2026, compared repeated ChatGPT answers to the same request. On average, 17.8% of all the products named across two answers appeared in both.

ChatGPT's app and its API, which a tracking tool can query instead of the app, averaged 12.0% overlap in the source domains they showed. The authors conclude that "A single response cannot characterise the recommendations or sources a system may expose."

We see the same in ChatGPT on a question about GEO firms that Omnius tracks for the US market. In nine completed runs from October 2 to 14, 2026, 10 different firms appeared in the top five, five different firms ranked first, and none appeared in all nine. An agency can at most shift the odds that a company is named, so any claim that visibility improved has to be measured against that noise.

When should a company hire a GEO agency, keep GEO in-house, or wait?

A company should hire a GEO agency for work its own team cannot deliver, keep GEO in-house when it has the people and access, and wait when prospects rarely use AI answers.

An internal team can correct pricing pages, change crawler permissions, and run answer tracking when it has the time. Outside help should fill a specific capacity gap, whether technical implementation, research, content, or publisher outreach.

The eight situations below each pair with one of five calls: fix a blocker first, run GEO in-house, hire for a specific gap, buy a scoped audit, or wait.

• Indexing or crawl access is broken. Fix SEO first, since engines can't cite pages they can't reach or index, and Search Console shows Google's gaps.
• Nobody can approve product facts quickly. Fix the approval path first, because every page and correction waits on sign-off, and a fintech's fee wording needs compliance review.
• The team has an SEO lead, writers, and a tracker. Run GEO in-house with any tracker that keeps the raw answer behind every score, such as Atomic AGI, the AI search tracking software the Omnius team builds.
• Own pages carry wrong facts. Fix them in-house, since Profound calls regular audits of owned media "an immediate lever for improving accuracy."
• Other websites carry wrong facts the team cannot reach. Hire for the outreach when nobody in-house knows those publishers and keep the scope tight, since in Profound's earned-media data such errors typically came from about four websites per affected brand.
• Rivals own the sources engines cite. Hire when competitors dominate the cited lists and articles and the company has no digital PR capacity.
• Prospects rarely use AI answers. Wait and monitor quarterly when discovery calls rarely mention an AI assistant.
• The budget covers an audit and a set of fixes. Buy that scope first with a retest date agreed in advance, and move to an ongoing engagement only when the remaining work justifies one.

Low AI referrals do not show low AI use, because a zero-click answer, one the reader acts on without visiting any website, can shape a shortlist. Conductor measured AI referrals at 1.08% of traffic across its US enterprise customers' domains in May to September 2025. The head of growth should therefore compare referral traffic and conversions with what prospects say on discovery calls.

Every branch starts with a rough baseline, a dated record of how AI answers treat the company before any work begins. The head of growth runs 20 to 50 prospect questions several times, signed out, in the engines prospects use, and dates the spreadsheet before the first agency call. Each row records whether the company is named, cited, or recommended and whether the answer gets its facts right. The same questions later go to every agency candidate.

Read more: In-house versus agency SEO

How to choose a GEO agency: Five things to evaluate

When hiring is the right call, a proposal is judged on measurement, proof, scope, fit, and terms, in that order. Measurement comes first because it makes the other four checkable. A case study or a reported change can be checked only when the same questions and baseline are used before and after. A single red flag can fail any of the five.

1. Measurement: Can the agency show a change on the company's own questions?

The measurement section should list the exact questions, the engines, the run frequency, the start date, and the testing conditions, such as country, language, and sign-in status. Any of the five left open lets a report drift to whatever moved. Visibility is then measured as a mention rate, the answers that name the company divided by the valid answers collected. Most questions should not name the company, such as best expense software for a 50-person finance team.

A single rank cannot describe answers that change from run to run. After SparkToro's test ran 12 mostly consumer prompts 2,961 times in late 2025, co-founder Rand Fishkin called "visibility % across dozens to hundreds of prompts run multiple times" a reasonable metric. Any tool that reports a "ranking position in AI," he added, is "full of baloney."

Branded questions flatter a score, and Peec AI reports mention-rate variance about three times higher for non-branded questions than for branded ones, so the two belong in separate scores.

Peec AI's analysis splits tracking into two jobs:

• Reading where a visibility score sits in a given month
• Spotting whether it has moved

For the first job, more distinct questions matter far more than repeat runs. On the same budget of runs, 50 questions run daily put a modeled monthly level within about ±6.1 points, while 10 questions run five times a day leave about ±13.2. For that first job, Peec's Jan Ehrlinspiel says that "increasing the amount of runs has nearly no impact."

Detecting a change on the same questions works differently, Peec explains, because each question is compared with itself over time, so more runs over more weeks reveal smaller moves. A proposal needs a broad question set and readings over many weeks, plus the smallest change it expects to detect.

Search Console's generative AI report shows how a website's pages perform in Google's AI features, and Bing's Citation Share gives its percentage of all citations shown for the web searches behind Microsoft's AI answers. Both measure pages and give a cross-check the agency does not collect itself. Microsoft says Citation Share "is designed as an observational metric."

Neither report shows why visibility moved, so the agency should log its own work and outside events. From October 8, 2026, for example, Profound saw ChatGPT drop the tracking tags from many of its links for about a week, so analytics could log those visits as direct traffic.

Red flag: a prompt count built on branded prompts, an average AI rank with no dated answers behind it, an API-only dashboard presented as what prospects see, or one score with no question list behind it.

What to do:

1. Send every candidate the questions from the rough baseline, drawn from sales calls, demo forms, and win-loss notes.
2. Score each measurement section on question source, branded split, engines, runs, baseline date, and a log of model releases. Then check that it counts mentions, citations, and recommendations separately, checks accuracy, and puts pipeline next to visibility.
3. Ask for an export of questions, raw answers, and citations.
4. Before the work starts, add a form question on how prospects found the company, with an AI follow-up, both saved to CRM fields, as our guide to AI search attribution describes.
5. Check that Search Console's generative AI report and Bing Webmaster Tools have data for the baseline weeks.

Read more: How to measure GEO success

2. Proof: Dated case studies with raw answers

Proof applies the same standard to past work. A case study or an AI visibility report can be checked when it gives dates, questions, engines, raw answers before and after, a change log, and the business signal.

Google's hiring guide suggests asking past clients whether the provider "produced positive results," and a GEO reference call can also ask which questions moved and what changed in pipeline. Google's June 5 guidance on third-party services adds that good advice "qualifies their claims as opinion based on data or experience" or cites official guidance. A screenshot of one answer is an anecdote, as the variance stage showed.

Red flag: an undated percentage, a rank position, a wall of logos, or a confidential case with nothing to inspect.

What to do:

1. Ask for one case with dated raw answers behind its chart, the number of answers, the change log, and any other changes, such as a model release.
2. Ask the reference what changed in the pipeline as well as in visibility.

Ask what the percentage measures. A case study headline that promises "+300% AI visibility" cannot be checked on its own, since it never says 300% of what, on which questions, in which engines, or over what period. A claim worth trusting names the question set, the engines, the dates, and the rate before and after. For example, a case study might report that the company was named in 18% of answers to 30 non-branded questions in ChatGPT and AI Mode in March 2026, and 41% in August.

3. Scope: Work on the website and off it, chosen from the baseline

More than half of the 300-plus people at SEO firms in Goodfirms' 2026 survey said they already sell AI search, so the GEO label alone says little about who can do the work. A useful scope ties each change to something the company can point to, such as a question from sales-call notes, a recurring support ticket, or a price ChatGPT gets wrong today.

Google's optimization guide, cited under access, narrows the scope for its own search. Structured data is optional, no special schema markup is needed, and a company can ignore content chunking and AI text files such as llms.txt.

A 2026 preprint studied 158,099 answers collected through the engines' APIs for two question sets written for client companies, with data help from a GEO services business. Pages whose titles and paragraphs matched the questions people ask were cited more, and that result held in a small B2B software market too.

In the larger consumer dataset, the study also replayed questions through the APIs, which can differ from the apps. There, whether ChatGPT's own search retrieved a page explained 32% of the differences in which pages of the same website got cited, and page features such as FAQ blocks explained 2.8%. The B2B data showed no statistically clear gap.

We treat entity authority, how consistently the web describes a company, as something earned on pages prospects already read. Google, too, says seeking inauthentic mentions "isn't as helpful as it might seem."

Red flag: a monthly post count with GEO in the title, tactics Google says its search can ignore sold as the plan's core, or paid placements sold as earned coverage.

What to do:

1. Ask which sources the engines cite for the company's questions today.
2. Check that each gap in the baseline has a named task and owner, including work on outside pages where the baseline shows a need.
3. Ask which parts are SEO the company already pays for and whether its current SEO provider could take on the rest.

4. Fit and delivery: Does the agency know the category, and what does it need?

Fit starts with evidence of work in the category. A head of growth can ask each candidate for dated work showing which sources AI engines cite for a comparable company's pricing. When Profound and a pricing consultant studied 7,675 AI answers about Cloud 100 companies, company-owned pricing pages appeared in 46% of answers. Vendr, a pricing data provider, appeared in 18.7% of runs. Naming familiar publishers is a starting point, and a relevant case with a reference shows the candidate's actual experience.

Fit also covers the delivery load: the products covered, pages to change, subject experts, access to the content management system (CMS), and approval times. For fintech work, the proposal should name who approves fee and license claims, the same approval path the hiring checklist says to fix first. An AI company that changes its plans often needs the fact audit rerun after each pricing change, since pricing and billing top the wrong claims, as the selection stage showed.

Red flag: no questions about the company's prospects, no named team, or a promise to handle everything.

What to do:

1. Ask who will do the work and who approves it on the company's side.
2. Ask for the first three months as dated deliverables with sign-off windows, such as a baseline and fact audit in weeks 1 and 2, then pricing pages in week 3.
3. Ask when the agency would advise against hiring it.

Read more: Choosing a B2B SEO agency

5. Terms and price: What the contract should say

Contract terms write the other four into the agreement, and they start with what an agency can honestly promise. Neither Google nor OpenAI sells a place in its organic answers: Google never accepts money to rank pages, and OpenAI says its ads "do not influence the answers ChatGPT gives."

Ads still appear next to organic answers, and Google says they may show "below and integrated into AI Mode responses," so a proposal should state what any guarantee covers. That could be paid ads, observed appearances on agreed questions, or a fee credit if a target is missed.

Published prices rise with company size and scope, and Clutch puts US enterprise programs at $150,000 to $500,000 or more a year. Neither GEO directory explains how its estimates were made, so quotes compare only on the same brief, with the same markets, deliverables, software, and exit terms priced.

Clutch's client reviews put the usual SEO project timeline at 12 months, so exit terms matter as much as price. Whito read 39 UK agency pricing pages against their own terms, and 19 of the 39 did not say how a client leaves.

What to do: Write these terms into the statement of work.

• A baseline in month one, before any work ships, shared as raw answers
• Deliverables itemized by month, each with its own price
• A 90-day review with success conditions named at signing
• The measured target, conditions, and remedy behind any guarantee
• Company rights to keep and export the question set, raw answers, reports, and content, with access terms for any licensed software
• Minimum term, notice period, and exit on the same page as the price

Red flag: an unconditional promise of organic citations, "placement" that turns out to be ads, a 12-month lock-in with no review, or data the company cannot export.

What questions should a company ask a GEO agency before hiring?

Those five factors turn into eight questions for every finalist, asked in the same order so the answers sit side by side. The last question tests whether hiring fits at all.

Each answer gets a pass, a partial, or a fail, and a proposal that fails on measurement or terms drops out whatever its price.

What does comparing 3 GEO proposals look like in practice?

In this worked example, a hypothetical Series B expense-management company sends 36 prospect questions to three candidates. It scores each proposal on the five factors, with the same pass, partial, or fail grades as the question table, and then reads the winner's first review against its baseline.

Proposal A guarantees placement in ChatGPT for 25 keywords and reports an average rank across 500 tracked prompts, and Proposal B sells eight blog posts a month optimized for AI. Proposal C tracks the 36 questions plus 6 from review websites in five engines weekly, adds a fact audit and two comparison pages, and needs sign-off within five working days.

When the company reruns the 42 questions at the 90-day review, the 18 category questions stay flat. On the 24 comparison questions, though, the company was named in 101 of 480 answers (21%) in the baseline month and 182 of 480 (38%) in the last month.

Proposal C's team reports the uncertainty for each engine and flags a ChatGPT model update in week 9 as a competing explanation, though the increase held for the four weeks after it. Prospects named an AI assistant in 4 of 37 new opportunities, so the company continues to the next review as a further test, with the question set unchanged.

How we at Omnius approach GEO in practice

We think of GEO as positioning a company inside a chain of systems from a prospect's question to the final answer. An engine interprets the question, retrieves candidate pages and companies by meaning and by exact wording, and narrows them by relevance, source quality, and evidence. A company can drop out at any step. Crawlers may not reach its pages, nothing may tie it to the category in question, or its claims may rest only on its own website.

Two beliefs follow from that chain. First, the company comes before the content. Engines retrieve from what the wider web associates with a company, so its entities and the evidence behind them decide whether it is considered at all. Second, answers are probabilistic. We measure visibility as a share of many repeated answers, start from the prompts closest to a purchase, and treat it as an intermediate variable on the way to pipeline and revenue.

In practice, an engagement runs in five phases, each with its own outputs:

1. Foundations and Day-0 baseline. Before any strategy, we map the company and its market: positioning, ideal customer profile (ICP), competitors, demand, pages, external sources, technical conditions, and past performance. Our own technology records how often ChatGPT, Perplexity, and AI Mode name, cite, and recommend the company for the prompts closest to a purchase, and which sources they cite.
2. Entity model. We map the company's entities across its website, profiles, registries, Knowledge Panels, competitors, and AI answers, then resolve conflicting data into one company information model. Each fact about its products, categories, use cases, customers, and integrations is marked as confirmed, inferred, or open. Entity coverage is benchmarked against competitors, and an external corroboration map shows which claims rest only on the company's own pages.
3. Strategy. The foundations and the entity model become a falsifiable strategic thesis and a small number of SEO and AI search pillars. The roadmap ranks each opportunity by intent, commercial value, evidence strength, probability of impact, and effort.
4. Execution on and off the website. Each priority entity gets one clear page, other pages are created, enriched, merged, or linked around it, and structured data ties people, products, and pages to one organization. Technical work opens access for AI crawlers such as GPTBot, ClaudeBot, and PerplexityBot. Off the website, the work goes to the lists, reviews, publications, and databases engines already cite, with digital PR and original data as their own workstreams.
5. Measurement and monitoring. The same technology connects Google, AI answers, citations, competitors, entity data, website behavior, technical signals, attribution, and conversions, benchmarked against more than 11,000 tracked domains. Its search ontology links brands, prompts, pages, citations, sources, and engines, so a change in an answer can be traced to the entity, page, or source behind it. Repeated sampling separates lasting patterns from randomness, and decay and competitor moves are flagged. Results are reported against the baseline next to organic sales-qualified leads (SQLs), pipeline, and revenue.

Behind these phases sits a methodology refined over more than a decade of B2B search work. It combines know-how from past clients, markets, experiments, and failures with documented standard operating procedures (SOPs) for inputs, decision rules, quality checks, and reviews, plus a knowledge base that keeps each company's context. One internal team runs every phase and owns the judgment, so each recommendation is tested against what happens after it ships.

That accumulated context is what raises the probability of results. For example, ANNA Money, a scale-up UK neobank for business owners and one of CNBC's top UK fintechs, saw clicks from AI search rise 164.44% and signups rise 60% between May and September 2025.

FAQ

Is a GEO agency the same as an AEO agency?

A GEO agency and an AEO agency usually sell the same work under different names. Both work on how AI answers mention, cite, and recommend a company, so the scope and the measurement show what a given agency delivers, whatever it calls itself.

What is the difference between a GEO agency and an SEO agency?

An SEO agency works to rank a company's pages in search results, while a GEO agency works on what AI answers say about the company, including mentions on third-party pages and the accuracy of facts. Many agencies offer both, so the scope and the measurement show which work is included.

How much does a GEO agency cost?

A GEO agency costs about $5,000 to $25,000 or more a month for a mid-market company in published estimates from Clutch and DesignRush. Prices rise with company size and scope, so quotes compare only when they price the same brief.

Enterprise programs cost more, and Clutch puts them at $150,000 to $500,000 or more a year.

Clutch also lists pilots of 8 to 12 weeks at $25,000 to $75,000. A pilot should name the deliverables and measured outcomes that decide whether the work continues.

Should a company check whether the agency itself shows up in ChatGPT?

An agency's own visibility in ChatGPT is a weak signal of client results, since its own pages can be among the sources cited. In a December 2025 study, Ahrefs found a recommended agency's own self-promotional list cited in 17.2% of ChatGPT answers to agency questions.

How long does GEO take to show results?

GEO results arrive at different speeds. We set the first review at about 90 days, to check shipped work, measured changes, and early commercial signals. Technical changes can show fast, since OpenAI says its systems can take about 24 hours to adjust after a robots.txt change. Google says recrawling a page can take several days to several months.

Methodology

This guide rests on platform documentation, independent and vendor studies, published price data, one Omnius observation, and the pages that rank for its primary query, read on October 20, 2026.

• Sources. The guide cites platform documentation, research and vendor studies, two GEO directories' price data, and Whito's contract research, all listed below. Guides by GEO firms were read for coverage only.
• Omnius observation. The observation covers nine completed ChatGPT runs (US) of one question about GEO firms that Omnius tracks, from October 2 to 14, 2026; the question and the firms are not named.
• Charts. The overlap chart reports Jaccard overlap from a preprint on physical products, the share of items named in either answer that appear in both. The margin chart extends Peec AI's published 50-question case with its rule that four times the questions halves the margin.
• Limits. The price ranges come from two GEO directories with different scopes, neither explains its method, and the chart leaves out DesignRush's entry-level tier for startups.
• Publishers' interests. Peec AI and Profound sell AI visibility tracking, Ahrefs and Conductor sell SEO software, the directories and review platforms have commercial ties to the firms they list, and one preprint had data help from a GEO services business. Omnius sells GEO services, and its team builds AI search tracking software.

Sources

The sources below are grouped by publisher.

• Google: Hiring an SEO, Optimizing for generative AI features, AI features and a website, Third-party SEO tools and advice, Search generative AI control, Generative AI performance reports, A new resource for optimizing, Documentation updates, Ask Google to recrawl, Ads in AI Overviews and AI Mode
• OpenAI: Our approach to advertising, Overview of OpenAI crawlers
• Microsoft: Generative AI and search ads, AI visibility insights in Bing Webmaster Tools
• Research: GEO: Generative Engine Optimization, Auditing AI-generated product recommendations, What drives AI citations?
• Studies and surveys: SparkToro on AI inconsistency, Peec AI on prompt tracking, Profound on citation sources, Profound on inaccurate claims, Profound on missing UTM tags, AI agents and pricing, Ahrefs on best lists, Ahrefs on brand correlation, Conductor benchmarks, G2 Answer Economy, TrustRadius survey
• Pricing and terms: Clutch GEO directory, Clutch SEO pricing, DesignRush GEO directory, Goodfirms SEO pricing, Whito on contract terms