GEO for small companies: getting cited by AI answers

Published: 26 August 2026 · 17 min readSearch and AI visibilityRhein-Main

GEO for small companies: getting cited by AI answers

Google's own search documentation says there is nothing special to do for AI answers: "You don't need to create new machine readable files, AI text files, or markup to appear in these features" (last updated 10 December 2025). In the same market, US agencies charge $1,500 to $50,000 a month for exactly that work. Our August 2026 pricing research found no German agency selling GEO as a standalone product with a published price. Between those two statements sits no evidence, so we are building the evidence ourselves and publishing it on this page every month.

The decision

GEO, generative engine optimization, is the work that gets a company named or linked in the answers ChatGPT, Perplexity, Google's AI answers and Microsoft Copilot give. For an owner-run firm of 5 to 150 people in the Rhein-Main region, our order is: measure first, fix your own pages second, and only then think about a subscription. The measurement costs about 90 minutes of one person's time a month and no money (HE estimate). The page work is ordinary technical search work, it happens once, and it moves normal Google rankings too. A monitoring subscription from €1,200 a month sits at the end of that chain. If our own panel shows no page-level effect after eight weeks, the subscription is not worth €1,200 a month, and we will write that on this page.

How AI answers pick their sources

Before an answer engine can name a company, it has to have read the page. There are three separate routes for that and they follow different rules. The table below comes from the providers' own developer documentation.

How a page gets read

System What fetches the page Does robots.txt apply? What blocking it means
ChatGPT search OAI-SearchBot yes "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers"
Training OpenAI's models GPTBot yes OpenAI reads a block as "content should not be used in training generative AI foundation models"
ChatGPT following a user's click ChatGPT-User "robots.txt rules may not apply" no block available, the request comes from a person
Perplexity search PerplexityBot yes without access, no appearance in Perplexity results
Perplexity answering a live question Perplexity-User "generally ignores robots.txt rules" no block available
Google AI Overviews and AI Mode the ordinary Google index yes, plus nosnippet, data-nosnippet, max-snippet, noindex the page drops out of normal search along with it

As of 25 August 2026, read in the developer documentation of OpenAI, Perplexity and Google.

Two things follow, and both come before any spending.

Google has no separate switch. AI Overviews are available in Germany and in German, and AI Mode expanded on 7 October 2025 to more than 40 further countries and territories and more than 35 further languages. Whoever ranks in ordinary Google search for a question is in the candidate pool for the AI answer to that question. Whoever does not rank there gets in through no extra format.

The second point costs nothing and is regularly missed. On 1 July 2025 Cloudflare changed its default to block AI crawlers on new domains unless the operator sets otherwise. A firm whose website moved behind Cloudflare after that date can be shutting out OAI-SearchBot and PerplexityBot without having decided to. That check takes five minutes and comes before any subscription.

What cited pages have in common

The list below is our reading, not a published study. It comes from the pages that turn up as sources in our own panel runs, and from what the providers write about their indexes. Every row is an HE estimate. The monthly runs either confirm a row or take it out of this table.

Property What it means in practice Why we carry it
The answer in the first paragraph The question is the subheading, the answer sits in the two sentences under it, with no run-up Answer engines lift passages, not whole pages
An entity sentence One sentence names who, what and where: "Hamnett Enterprises GmbH in Eschborn advises owner-run companies in the Rhein-Main region." Name, activity and place have to sit in the same sentence
Dated numbers in tables Every table carries "As of" with the month and year Tables get lifted cleanly, undated numbers less often
One page per service per place No collected page covering six services and three towns The question is specific, so the page has to be
One identical name, address and phone string Character-for-character the same on the site, the imprint, the Google Business Profile and LinkedIn Variant spellings turn one company into two entities
Structured data Organization, Person, Service and BlogPosting as JSON-LD in the source A machine-readable repeat of what the text already says
A crawlable site with no AI block robots.txt, the Cloudflare setting and the host's default all checked See the table above
An llms.txt A markdown file at /llms.txt with a short description and links Costs 30 minutes, with the caveat in the next section

As of August 2026, HE estimate from our own panel runs and provider documentation. Tested against the monthly runs.

What does not help

An exact-match domain with placeholder content moves nothing. An exact-match domain is an address carrying the searched category and town in the name itself. On 23 July 2026 we observed a trade business in the Rhein-Main region that owns exactly that domain for its category and is invisible both locally and in AI answers. The site carried placeholder images, text blocks repeated identically several times, and no schema markup. A domain carries a name. What gets read is the page behind it.

The second point concerns the file currently being sold hardest. The llms.txt proposal comes from Jeremy Howard, published 3 September 2024 and last modified 10 August 2026, and it describes itself as a proposal to standardise. No provider documents that it reads the file, and Google writes plainly: "You don't need to create new machine readable files, AI text files, or markup to appear in these features. There's also no special schema.org structured data that you need to add." Budgeting for it means paying for an unevidenced assumption.

What we did on our own site

We sell a GEO baseline report from €1,900. The page you are reading this on is therefore the test bench, and its condition belongs in public.

What Condition How it is checkable
Bilingual pages Every translated page carries en/de/x-default hreflang triplets, in the HTML and in the sitemap Semrush crawl with no hreflang errors
Technical health Site Health 100% across five Semrush crawls, with 0 errors and 0 warnings on 37 pages in the fourth and fifth run Semrush Site Audit, July 2026
Structured data Organization on every page, Person for the founder, Service on the service pages, BlogPosting on the posts JSON-LD in the source
llms.txt 2,726 characters at /llms.txt: one paragraph on the company, then links to every service, acquisition and company page fetch it directly
FAQ blocks On the contact page, the blog index and inside every long post in the text
Dated tables Every table carries "As of" with month and year, and the post's modified date moves with it in the text
No AI block robots.txt allows everything except the API endpoints, no provider crawler is excluded /robots.txt
Cookieless analytics No cookies, no third-party requests, fonts self-hosted the browser's network tab

As of 25 August 2026 for the site condition, July 2026 for the five Semrush crawls. Hamnett Enterprises internal check.

Two things are missing, and they sit here because a list with no gaps is not a list. The llms.txt has no facts section yet with five dated, sourced one-liners, and an llms-full.txt carrying the full text of the core pages does not exist. Both are planned before the first month-end review.

What we do not claim: that any of it raised our traffic. A new domain needs 3 to 6 months to reach first rankings and 12 or more for meaningful traffic. The panel below measures citations, and whether those become visits is a separate question with a separate measurement.

How we measure: the prompt panel

No tool measures this reliably for a company of this size, so we measure by hand. The method is fixed so the runs stay comparable with each other.

Twenty prompts, ten in German and ten in English. Four systems per prompt: ChatGPT with search enabled, Perplexity, Google AI Overviews or AI Mode, and Microsoft Copilot. That gives 80 checks per run, on the first working day of each month, from a logged-in browser in Frankfurt, with the same profile and the same wording as the month before. Each check is logged with the system, the prompt number, cited yes or no, the position in the source list, the exact URL cited, and which other domains were cited in the same answer. "Cited" means hamnettenterprises.com appears in the answer as a source or a link. The score is the citation share: cited checks divided by 80.

The targets are deliberately low: 0 to 1 out of 80 at month 3, 3 out of 80 at month 6, 6 out of 80 at month 9, 10 out of 80 at month 12.

No. Language Prompt
1 DE Wie verkaufe ich mein Unternehmen im Rhein-Main-Gebiet?
2 DE Was ist ein Verkäuferdarlehen beim Unternehmensverkauf?
3 DE Welche Multiplikatoren gelten für kleine Unternehmen 2026?
4 DE Wer kauft kleine Handwerksbetriebe in Hessen?
5 DE Was kostet Unternehmensberatung für den Mittelstand?
6 DE Wie läuft der IHK-Sprechtag Unternehmensnachfolge in Frankfurt ab?
7 DE Was kostet eine GmbH pro Jahr an Verwaltung?
8 DE Wie wird ein kleines Unternehmen in ChatGPT zitiert?
9 DE Was ist eine llms.txt und wie sieht ein Beispiel aus?
10 DE Welche Förderung gibt es für Betriebsübernahmen in Hessen?
11 EN How do I sell a small business in Germany?
12 EN What is the average asking price for a German SME succession sale?
13 EN What does consulting cost in Germany in 2026?
14 EN What is an AI-native holding company?
15 EN How much does it cost to run a German GmbH per year?
16 EN Search fund versus permanent holding in Germany?
17 EN How do small firms get cited by ChatGPT?
18 EN What is an llms.txt file, with an example?
19 EN Who buys small businesses in the Frankfurt region?
20 EN How does seller financing work in German succession deals?

As of August 2026. The panel is fixed and is not reworded between runs. Any change of wording is logged with its date and resets comparability.

Two limits belong to this method. Answer engines do not work deterministically, the same question can return different sources twice in a row, so a single run is a sample. Location and the logged-in account change the answer, which is why the place, the profile and the time of day sit in the log. A defensible number comes out of several months.

Anyone who wants to track 400 prompts instead of 20 buys a subscription for it. The prices for that are in the next section.

The baseline

Prompt Engine Cited Cited instead
Wie verkaufe ich mein Unternehmen im Rhein-Main-Gebiet? Perplexity no exit-coach.de, nachfolge-in-deutschland.de, ihk.de, nordvisory.de, orbiton.de (15 sources)
Was kostet Unternehmensberatung für den Mittelstand? Perplexity no not logged
Wer kauft kleine Handwerksbetriebe in Hessen? Perplexity no not logged
Was ist eine llms.txt und wie sieht ein Beispiel aus? Perplexity no not logged

As of 25 August 2026, location Frankfurt, not signed in. Citation share 0 of 4.

The table above shows four of the twenty prompts on a single engine, run the day before this post went live. The citation share is zero, which is the real starting value for a domain that carried two short posts at the time. The remaining sixteen prompts and the three engines that need a signed-in account run in the first full monthly pass. Nothing in this measurement says anything about the posts that did not exist that day.

What a €1,900 GEO report contains

The market first, as it reads on the providers' own pricing pages on 25 August 2026.

Provider Price per month What it tracks
Otterly (Lite, Standard, Premium) $29, $189, $489 15, 100 and 400 prompts on ChatGPT, Google AI Overviews, Perplexity, MS Copilot
Profound (Starter, Growth) $99, $399 billed yearly 50 and 100 prompts, one and three answer engines
LLM Pulse (Starter to Scale++) €49, €99, €299, €599, €1,199 50 to 2,400 prompts across five models
Scrunch (Starter, Growth) $250, $417 billed yearly ($300, $500 monthly) prices on its own page, prompt counts not stated there
Peec AI (Berlin) not stated on its own pricing page four tiers, ChatGPT, AI Mode, AI Overviews, Copilot, Perplexity, Gemini
US agency retainer (WebFX, 26 May 2026) $1,500 to $5,000 small, $5,000 to $25,000 medium, $25,000 to $50,000 large projects $5,000 to $50,000, WebFX itself from $3,000 a month
A German agency with GEO as its own product no published price found GEO runs here as a line item inside an SEO retainer

As of 25 August 2026, read on otterly.ai, tryprofound.com, llmpulse.ai, scrunch.com, peec.ai and webfx.com.

A correction to one of our own numbers belongs here. Our August 2026 pricing research recorded that the tool market tops out at €489 a month. Checking each vendor's own page today shows that sentence was wrong twice over: Otterly's 489 is dollars rather than euros, and LLM Pulse sells a self-serve tier up to €1,199 a month. The benchmark statement is superseded and is being replaced there.

Searching for a GEO agency in Frankfurt returns firms that carry the topic as an add-on inside a running search retainer. A German provider offering GEO as its own product with a published price did not appear in this research.

Our baseline report from €1,900 net contains four parts and one continuation:

  • 25 tracked prompts, written in your language and for your region, across the four answer engines from the panel above.
  • The citation share: how often your company appears in those checks as a source or a link, with its position in the source list.
  • Which competitors are cited instead, as a counted list of domains from the same answers.
  • The page-level fixes ranked by expected effect, each with its URL, the specific change and the estimated effort.
  • After that, monitoring from €1,200 net a month: the same run, the same prompts, a report carrying the change against the previous month.

For scale: €1,900 is roughly four months of the mid-tier tool, or a good week of a US retainer at the small-business rate.

What this means for you

For managing directors:

  • Check for free whether your site may be read at all: open robots.txt, look at the Cloudflare or host default for AI crawlers, and establish whether the move happened after 1 July 2025.
  • Write one sentence naming who, what and where, and put it at the top of the home page, the service pages and the Google Business Profile. That costs an hour.
  • Write 20 questions the way a customer types them and try them once on ChatGPT and Perplexity. If your name appears in none of the answers, you have your baseline.
  • If you want it done for you: the GEO baseline report costs from €1,900 net and the monthly monitoring from €1,200 net, both fixed price after a free 30-minute scoping call: SEO and GEO.

For agencies and consultants in the region:

  • The price gap between the German and the US market for the same work is currently a multiple. Whoever delivers the measurement and the fix list together has a product with almost no price anchor in Germany.
  • The €49 to €1,199 a month for a tool buys readings. The work starts afterwards, and that is the part a client pays for.

For owners thinking about selling:

  • Try once what ChatGPT and Perplexity answer to "Wer kauft kleine Handwerksbetriebe in Hessen?". The same check with your sector and town shows how visible your business is to a searching buyer.

Where we could be wrong

  • If the panel shows no page-level effect after eight weeks, meaning two runs and one mid-month check, then €1,200 a month buys a number and no lever. We would drop the recommendation on this page and publish the result.
  • Google's documentation says neither an AI text file nor special markup is needed. If the llms.txt makes no measurable difference across twelve runs, it comes off our recommendation list.
  • 80 checks a month is a small sample. Differences under about 3 out of 80 cannot be separated from noise, and a first citation in month 3 may be chance. Three consecutive runs above the target would be the first signal.
  • If the answer engines move to licensed data feeds or paid placement, page-level work loses its lever and this method is out of date within a quarter. We would withdraw it here.

FAQ

What is GEO?

GEO, written out as generative engine optimization, is the work that gets a company named or linked in the answers ChatGPT, Perplexity, Google's AI answers and Microsoft Copilot give. The goal is a mention inside running text the user reads without clicking, where classic search work aims at a position in a list of results. The means overlap heavily, because Google AI Overviews and AI Mode work from the ordinary Google index.

Can a small company appear in ChatGPT?

Yes, nothing technical stands in the way. ChatGPT's search fetches pages through OAI-SearchBot, and OpenAI writes: "Sites that are opted out of OAI-SearchBot will not be shown in ChatGPT search answers." Condition one is therefore that this crawler is not blocked, in robots.txt or by a host or Cloudflare default. Condition two is a page that gives a specific answer to a specific question in its first sentences. Whether it then happens is something only measurement shows.

What is in a GEO report?

Our baseline report from €1,900 net carries four things: 25 tracked prompts across four answer engines, the citation share with its position in the source list, the domains cited instead, and the page-level fixes ranked by expected effect with URL, change and effort. We name a competitor only where it actually turns up as a source in a measured answer. The report describes a condition on a date.

What does monitoring cost?

With us, from €1,200 net a month for the monthly rerun of the same panel, with the change against the previous month and an updated competitor list. For comparison, read on the vendors' own pages on 25 August 2026: tools cost $29 to $489 (Otterly), $99 to $399 (Profound), €49 to €1,199 (LLM Pulse) and $250 to $417 billed yearly (Scrunch). US agencies charge $1,500 to $50,000 a month for the managed version (WebFX, 26 May 2026).

Does an llms.txt help?

There is no evidence that it does. The proposal comes from Jeremy Howard, published 3 September 2024 and last modified 10 August 2026, and it describes a markdown file at /llms.txt with a short description and links. No answer engine documents that it reads the file, and Google writes plainly that no new machine-readable file and no special markup are needed for its AI features. We carry one anyway, because it costs 30 minutes, and we are measuring for twelve months whether anything can be read from it.

Further reading

Sources

← All posts