A GEO agency works to get your business named and cited when people ask ChatGPT, Claude, Gemini and Perplexity questions about your field: it makes sure AI crawlers can reach your site, publishes pages that answer your buyers' questions, and measures how often the assistants actually name you. Choose one by the evidence it can produce, meaning dated, repeated measurements of assistants' answers to a fixed set of questions, and walk away from anyone who guarantees citations or promises traffic.
What does a GEO agency actually do?
The term generative engine optimization comes from a paper by Aggarwal et al., presented at KDD 2024, which found in controlled experiments that certain content changes could raise a source's visibility in generative engine responses by up to 40%. That is a benchmark result, not a forecast for your business, but it captures the idea: assistants pick sources, and some pages are easier to pick than others.
In client work, the job splits into three parts.
Crawler access and indexing
If crawlers cannot fetch your pages, nothing else matters. A competent agency checks robots.txt, sitemaps, firewall and CDN rules, and whether Bing and Google have actually indexed your pages, and it submits new URLs through IndexNow.
This is not a formality. On our own site, a robots.txt rule meant to hide one redirect page (/b) was written as a prefix. Because robots.txt rules match prefixes, from 11 August to 21 October 2026 it also closed /blog, /business and every article in our business blog to all crawlers. Bing Webmaster Tools listed the business page as 'discovered but not crawled' the whole time, and Google had none of those pages indexed. We changed the rule to /b$ on 21 October 2026. If it can happen to a team that does this every day, it can happen on your site; this guide to blocked AI crawlers shows how to check.
The crawlers are busy when the door is open. Across the retained nginx logs of an edtech platform we work with, pickaclass.com, from 28 July to 11 August 2026, ClaudeBot made 19,272,125 requests, GPTBot 2,744,985 and Googlebot 52,292.
Pages that answer buyer questions
The second part is publishing, on your own domain, pages that answer what your buyers type into assistants: what something costs, how to choose, what can go wrong, who to hire.
The amount of useful content matters more than polished markup. In our data, a site with complete markup (server rendering, Article, FAQPage and Organization JSON-LD, a sitemap) but only 75 URLs was still absent from answers on its own article's exact topic, while a site with about 190,000 crawlable URLs was crawled constantly. Structured data is worth having, and Google says to treat Google Search Central, not schema.org, as definitive for what Google uses. It is not a substitute for pages worth citing.
Measurement
The third part is asking the assistants your buyers' questions on a schedule and counting the answers that name you and the sites they cite. Many agencies are thin here, yet it is the only way to know whether the first two parts are working.
Why can't anyone guarantee that an assistant will name you?
Because the assistants decide, and they decide differently each time.
- Answers vary. Asked the same question three times, an assistant often names different businesses. In a test on 23 September 2026, two runs of the same question named two different sets of firms.
- Many answers involve no search. In that test of 20 German business questions, the model answered 10 without searching and cited nobody. Where it searched, it cited primary sources (laws, ministries, the OECD, large accounting firms) on 9 of 10. Firms were named only on 'who should I hire' and 'how to choose' questions.
- Indexed is not found. On 19 October 2026 we asked assistants for the exact titles of 18 articles, each older than three weeks, on three client sites. ChatGPT found 1 of 18; Claude found 7 of 18. Google had indexed 29 of 30 articles of one of those sites.
When OpenAI's models do use web search, the API documentation says responses include inline citations with the URL and title of each source. The catch is the 'when'. Researchers see the same opacity: an ICML 2026 position paper calls for disclosure standards and black-box auditing of AI answer engines, because outsiders cannot see how sources get chosen.
What are the red flags when hiring a GEO agency?
- Guaranteed citations or 'recommended by ChatGPT'. Nobody controls the model. A guarantee means the agency either misunderstands the work or expects you not to check.
- Traffic promises. In about two weeks of logs from three of our course websites (roughly 1.2 million requests in the two busiest logs), human visitors arriving from AI assistants were: Copilot 1, every other assistant 0. Assistants mostly answer in place. The value shows up as being named, and in live retrievals: in the same fortnight the edtech platform received 1,918 fetches from ChatGPT-User, OAI-SearchBot and Perplexity-User on content pages. A click forecast suggests nobody looked at logs.
- Screenshots as the main evidence. One screenshot is one answer out of many that vary.
- Markup or llms.txt sold as the fix. Useful hygiene, not the work.
- No crawl audit in the first weeks. If nobody opens your robots.txt, mistakes like ours stay hidden.
- Reports that change the questions. Weeks are only comparable on the same questions and the same number of asks.
- Long lock-in contracts in a field where outcomes are uncertain.
What questions should you ask before signing?
- Which questions will you track, and can I see and edit the list?
- Which assistants, how many asks per question, and how often?
- How do you count a mention when answers vary between asks?
- Will I see which sites the assistants cite instead of me?
- What will you check and fix technically in the first month?
- Is the content published on my domain or yours?
- What is the contract term, and how do I leave?
- Have you measured your own visibility, and what did you find?
The last question is telling. On 21 October 2026 we put 20 English and 20 German questions about getting a business into AI answers to ChatGPT, Claude, Gemini and Perplexity, three times each. We were named in 0 of 160 answers (full write-up here). An agency that reports its own numbers honestly is more likely to report yours honestly.
What proof should you demand?
Measured answers, not screenshots.
| Aspect | Screenshot | Measured answers |
|---|---|---|
| What it shows | One answer at one moment | Every answer to a fixed question set, each week |
| Variation | Hidden; the best run gets shown | Counted, e.g. named in any of three asks |
| Comparability | None between months | Same questions, same number of asks, week over week |
| Sources | Usually cropped out | Lists the sites each assistant cited |
| Reproducible | Rarely | Yes, with the question list and dates |
Measuring is cheap, so there is no excuse to skip it. At API prices in October 2026, one question asked once with web search costs us about $0.069 with Claude and $0.0013 with Perplexity; a report of 20 questions x 4 assistants x 3 asks costs about $5.40.
Demand a baseline before work starts. A real one looks like this: for an orthopaedic practice in Bonn on 20 October 2026, the four assistants named the practice in 12 of 80 answers to 20 patient questions (Claude 5 of 20, Gemini 4, ChatGPT 2, Perplexity 1). The directory jameda.de was cited in 12 answers, and competing practices' own websites were the most cited sources. That tells you where to compete.
Client stories are fine when labelled. Ours, as reported by the clients in October 2026: FreightSizer, a B2B SaaS for freight dimensioning, got 2 new B2B customers in one month, both from ChatGPT; appslift, an app store optimisation service, gets 2 to 5 leads a month from AI assistants. Treat such claims as self-reported and pair them with measured data.
One trap: asking ChatGPT which GEO agency to hire and trusting the list. In our 21 October 2026 measurement, the most cited sources for GEO questions were the blogs of GEO agencies and tool vendors themselves. Being named there shows an agency publishes well about itself, which is real skill, but it is not evidence of results for clients like you.
What does it cost, and what should you do next?
Prices vary widely; this breakdown of monthly GEO costs covers the range. ItGrows for Business costs $499 a month (plus a one-time $499 setup, free with 3 months or a year): it publishes articles on your domain, fixes crawler access and indexing, and every week puts your buyer questions to the four assistants three times each, counting answers that name you and the sites cited. Like anyone honest in this field, we cannot guarantee that an assistant will name you.
Before talking to any agency, get your own baseline. The free check generates five buyer questions from your homepage and shows in 25 to 40 seconds whether ChatGPT and Perplexity name you. You can also check by hand with this guide. Then bring that baseline to every sales call and ask exactly how they would move it.
Frequently Asked Questions
How long before a GEO agency shows results?
Nobody can promise a date, and an agency that does is guessing. In our weekly measurement of 20 October 2026, Perplexity cited an article we had published on a client's site, a German company-formation advisory, which is the kind of early signal to watch for. Judge progress on the same questions, asked the same number of times, over months.
Is a GEO agency different from an SEO agency?
The crawl, indexing and content work overlaps. The difference is what gets measured: answers from assistants rather than search rankings. In our 19 October 2026 test, being in Google's index did not mean an assistant could find the page.
Can I do GEO myself instead of hiring an agency?
Yes. You can audit robots.txt and sitemaps, publish answer pages and ask the assistants your buyers' questions by hand. The hard parts are consistency, asking each question several times, and keeping the question set fixed so weeks are comparable.
Which tracking tools do assistants name for GEO?
In our 21 October 2026 measurement of English answers, assistants named Profound in 12 answers, Semrush in 9, Otterly.AI and Peec AI in 8 each, and Ahrefs in 7. Whatever tool or agency you use, ask to see the raw questions, dates and answer counts behind any report.