How to choose a GEO agency: the claims to test before you buy AI search work
Generative engine optimization is new enough that almost any claim sounds plausible in a proposal. The engines publish how they find pages, and one peer-reviewed study measured what changes their answers. Test every proposal against both before you sign.
The short version
- Google says there are no additional requirements or special optimizations needed to appear in its AI Overviews or AI Mode. A page must be indexed and eligible for a snippet. Treat any proposal built on secret AI markup with care.
- The one large peer-reviewed study found that citing sources, adding quotations and adding statistics raised a source’s visibility in AI answers by 30 to 40 percent on its main measure. Keyword stuffing did little, and on Perplexity it did worse than doing nothing.
- Training crawlers and search crawlers are different. Blocking OpenAI’s training crawler does not remove you from ChatGPT search; blocking its search crawler does. An agency that cannot explain the difference should not be touching your robots.txt.
- Nobody can guarantee what an assistant says about you. What can be measured is whether your pages are crawlable, quotable and corroborated, and how often you are named for the questions your buyers ask.
- Judge a proposal on what it will measure, from which baseline, and how that connects to enquiries. A visibility score with no method behind it is not a result.
Summarise with AI
ChatGPT
Claude
PerplexityHow to choose a GEO agency
To choose a GEO agency, ask each one for four things in writing: the baseline it will measure from, the pages it will change and what it will add to each, which AI crawlers it will allow and block, and how it will connect the work to enquiries. Then test its claims against what the engines publish about themselves and what independent research has measured. Guarantees, secret AI markup and keyword work are the warning signs.
This guide is for marketing leads who have decided AI answers matter to their pipeline and are now comparing agencies to do the work. If you are still deciding whether it matters, start with what your buyers actually ask assistants about your category, and our GEO page explains what gets a business named in those answers.
The proposals you will receive describe the same work in very different words: generative engine optimization, answer engine optimization, AI search optimization, LLM SEO. Underneath, the work is the same three things. Make your pages findable by the crawlers assistants rely on, make them quotable when an assistant reads them, and make the facts about you consistent on the third-party sites assistants also read.
Google is unusually direct about its own AI features. Its guidance for site owners says there are no additional requirements to appear in AI Overviews or AI Mode, nor other special optimizations necessary. A page needs to be indexed and eligible to be shown with a snippet, and the existing SEO fundamentals still apply. That does not make GEO a fiction. It means the work is ordinary search work done with a new reader in mind, and a proposal that depends on hidden techniques deserves questions.
One disclosure before the rest. We sell GEO services, so we are one of the agencies this article is about. The tests below are the ones we would want a buyer to apply to us, and the evidence comes from the engines’ own documentation and independent research.
What the research says changes an AI answer
Most claims about GEO come from agencies. The most rigorous independent evidence is a paper by researchers at Princeton University and IIT Delhi, published at KDD 2024. They built a benchmark of 10,000 queries, rewrote source pages in nine ways, and measured how much of each generated answer came from each source.
| Change to the page | What happened to its visibility in AI answers |
|---|---|
| Citing sources, adding quotations, adding statistics | The top methods. A relative improvement of 30 to 40 percent on the study’s main measure and 15 to 30 percent on its subjective measure |
| Improving fluency and readability | Also a significant improvement, of 15 to 30 percent |
| A more persuasive, authoritative tone | No significant improvement |
| Keyword stuffing | Little to no improvement on the study’s engine, and 10 percent worse than doing nothing on Perplexity |
| Pages ranked lower in search | Benefited more. Citing sources raised visibility by 115.1 percent for the source ranked fifth, while the top-ranked source lost 30.3 percent on average when every source was optimized |
On Perplexity, a real engine with a large user base, the same methods held up: adding quotations performed best on the study’s main measure, with a 22 percent improvement, and the improvements reached 37 percent on the other.
Three limits matter before you use these numbers in a decision. The study measured how much of an answer a source contributed, not clicks, enquiries or revenue. The benchmark was built by the researchers and tested on their own engine and on Perplexity as it behaved in 2023 and 2024, and engines change. And the authors found the effect varied by domain, which is the opposite of a universal recipe. Read it as a direction: evidence and specifics help, padding and posture do not.
That direction is useful for judging proposals. Work that adds sources, figures with their origins, quotable definitions and clear answers is consistent with the evidence. Work that adds keywords or a more confident tone is not.
The crawler question every proposal should answer
Assistants find pages through crawlers, and most companies now run more than one: one that collects content for training models, one that builds a search index, and one that fetches a page when a user asks. Blocking each one has a different effect. Your robots.txt decides which ones may read your site, so a GEO agency will either touch it or tell you to.
| Crawler | What it is for | What blocking it does |
|---|---|---|
| OAI-SearchBot (OpenAI) | Surfaces websites in ChatGPT’s search features | Your site will not be shown in ChatGPT search answers, though it can still appear as a navigational link |
| GPTBot (OpenAI) | Crawls content that may be used to train OpenAI’s models | Signals your content should not be used for training. It is not the search crawler |
| ChatGPT-User (OpenAI) | Visits a page when a ChatGPT user’s request needs it | User-initiated, so robots.txt rules may not apply; it plays no role in search visibility |
| PerplexityBot | Surfaces and links websites in Perplexity’s search results; not used for training | Your site drops out of Perplexity’s results |
| Claude-SearchBot (Anthropic) | Improves search result quality for Claude users | May reduce your visibility and accuracy in Claude’s search results |
| ClaudeBot (Anthropic) | Collects content that could contribute to model training | Signals your content should be excluded from training |
| Google-Extended | Controls whether Google may use your content to train future Gemini models | Google says it does not affect inclusion in Google Search and is not a ranking signal |
The sources are OpenAI’s crawler documentation, Perplexity’s, Anthropic’s and Google’s list of its crawlers.
The practical point: whether to let your content train models is a business and legal decision for you to make. Whether to let search crawlers in is a visibility decision, and blocking them removes you from the answers you are paying an agency to get you into. A competent proposal separates the two, names the crawlers it will allow, and says why. One that offers to unblock all AI bots, or to block them all, has not read the documentation.
Claims to test in a GEO proposal
Put each of these to every agency on your shortlist, including us. The answer matters more than the claim.
“We guarantee you will be recommended by ChatGPT”
Nobody controls what an assistant says. Ask what will be measured instead, and how often.
“We add special markup for AI”
Google says no special optimization is needed for its AI features. Ask which change, on which page, and what evidence says it matters.
“We optimise your keywords for AI”
Keyword stuffing did little in the GEO study and did worse than nothing on Perplexity. Ask what will be added to each page besides keywords.
“Your AI visibility score is low”
Ask which questions were asked, on which assistants, how many times, and whether the check can be repeated to show change.
“We will fix your robots.txt for AI”
Ask which crawlers they will allow and block, and whether they separate training crawlers from search crawlers.
“We will make your content more authoritative”
Tone alone showed no significant improvement in the study. Ask whether they will add sources, figures and quotations, which did.
“We will get you onto the lists AI reads”
Third-party sources matter. Ask which sites, how you get there, and whether any placement is paid, so you can judge what it is worth.
“You will see results in weeks”
Ask for the baseline they will measure from, and what result would show the work is not working.
Is this GEO proposal worth buying?
Tick what the proposal in front of you does. Score each agency separately.
What a GEO engagement should measure
AI visibility is harder to measure than rankings, which is exactly why a proposal should say how it will be done. Four measures are worth agreeing before the work starts:
- A fixed set of the questions your buyers ask, put to named assistants on a schedule, recording whether you are named, cited or linked. The same set every time, so change means something.
- Which of your pages are cited, so the work can follow what assistants actually use.
- Search performance in Search Console for the pages being changed, because Google’s AI features draw on the same index as its results.
- Enquiries that arrive from AI answers. The simplest instrument is a question on your enquiry form asking how the buyer found you, with the assistants listed as answers. It is crude, and it is the measure closest to revenue.
Agree what would count as failure too. If the baseline questions show no change after a fair period, a good agency will say so and change the plan rather than change the measure.
Where GEO sits next to the rest of your search work
Because assistants and AI features draw heavily on search indexes, GEO rarely works on a site with weak search foundations. If pages are not indexed, are slow or say little that is specific, fix that first, and B2B SEO is the place to start. If the site itself needs rebuilding, the website redesign checklist covers what not to lose on the way.
And if buyers in your category do not ask assistants about suppliers yet, paid search may produce enquiries sooner. PPC for SaaS covers how to make an ad account learn from the right leads.
Questions about choosing a GEO agency
How do I choose a GEO agency?
Ask each agency for the baseline it will measure from, the pages it will change and what it will add to each, how it treats AI crawlers, which third-party sources it will work on, and how it connects the work to enquiries. Then test its claims against the evidence: guarantees, secret markup and keyword work are warning signs.
Can a GEO agency guarantee my company will appear in ChatGPT?
No. Nobody controls what an assistant says. An agency can make your pages crawlable, quotable and corroborated, and measure how often you are named for a fixed set of questions. A guarantee beyond that is a sales line.
Is GEO different from SEO?
It overlaps heavily. Google says its AI features need no special optimization beyond being indexed and eligible for a snippet. What GEO adds is writing for an assistant that quotes rather than ranks, and work on the third-party sites assistants also read.
Should I block AI crawlers?
Separate the two kinds. Blocking training crawlers, such as GPTBot, ClaudeBot or Google-Extended, keeps your content out of model training and, by their owners’ accounts, does not remove you from search. Blocking search crawlers, such as OAI-SearchBot, PerplexityBot or Claude-SearchBot, removes or reduces you in those assistants’ answers.
How is GEO measured?
By a fixed set of buyer questions put to named assistants on a schedule, the pages that get cited, search performance for those pages, and enquiries that say they came from an AI answer.
Method and sources
The study and the crawler documentation below we read at source in October 2026; the study’s figures are quoted with its limits in the text. The engines change their crawlers and features often, so check each company’s page before acting on the table. The claims to test and the measures are our judgement from doing GEO and SEO work, written as judgement rather than as data.
- GEO: Generative Engine OptimizationAggarwal, Murahari, Rajpurohit, Kalyan, Narasimhan and Deshpande, KDD 2024. A 10,000-query benchmark; citing sources, quotations and statistics improved visibility by 30 to 40 percent; keyword stuffing did little and was 10 percent worse on Perplexity; lower-ranked sources gained most
- AI features and your websiteGoogle Search Central. No additional requirements or special optimizations for AI Overviews and AI Mode; pages must be indexed and snippet-eligible
- Overview of OpenAI crawlersOpenAI. OAI-SearchBot, GPTBot and ChatGPT-User, and what disallowing each does
- Perplexity crawlersPerplexity. PerplexityBot and Perplexity-User
- Does Anthropic crawl data from the web, and how can site owners block the crawler?Anthropic. ClaudeBot, Claude-User and Claude-SearchBot
- Google’s common crawlersGoogle Search Central. Google-Extended does not affect inclusion in Google Search and is not a ranking signal
