The 10 most common GEO mistakes, and how to avoid each one

Mistakes that stop the assistants finding you

1. A bot wall nobody chose

In our audit (40 trades, 90 metros, September 2026), 7.2% of the sites served our crawler a challenge page instead of the site. Nobody at those businesses decided to hide from the assistants; a host's default, a CDN setting or a security plugin did it for them. The fix is a question to whoever runs that layer: which crawlers do you let through, by name? Then fetch your homepage from outside your own network and confirm a page comes back.

2. Blocking the wrong crawler

The tokens do different jobs. OpenAI's crawler documentation separates GPTBot, which collects training data, from OAI-SearchBot, which indexes pages for ChatGPT search; blocking the first does not remove a site from search answers, and blocking the second does. Google's Google-Extended token governs Gemini's training and grounding without touching Search, while Googlebot feeds Search itself. In the audit, 3.2% of sites disallowed AI crawlers on purpose. Decide each token on its own, and keep the answer-side ones open; the crawler access check reads your file for all of them.

3. Facts locked in a PDF, a picture or a script

2.1% of the audited sites kept their real content, the price list or the service menu, in a PDF, and 3.0% delivered their text only through JavaScript, which a crawler that runs no scripts receives as an empty page. A photo of the menu board or a price list set in an image states nothing to a parser at all. Put the facts in the page's HTML as text and keep the PDF, if you like it, as the printable copy. The PDF trap covers files in detail.

4. Three versions of the business

43.1% of the sites told a different story about who and where they were across their own homepage, contact page and about page. 36.9% showed more than one phone number, and 34.1% gave no postal address a machine could parse. A reader matching those pages against a listing has to guess whether it is looking at one business or three. Choose one form of the name, one address and one main number, write them on your own site first, and copy them outward, character for character.

5. Markup that describes the website, not the business

17.8% of the sites carried no schema.org markup at all. Most of the rest described the page rather than the business, with WebSite, WebPage and BreadcrumbList nodes a page builder added by itself, and LocalBusiness appeared on roughly a quarter of the sites. A further 13.7% had business markup that was missing the fields that make it usable. Mark up the business with the most specific LocalBusiness type that fits, fill in the name, address, telephone and hours, and repeat only what the page shows; structured data for AI has the block.

Mistakes that stop an answer recommending you

6. Questions asked in headings and never answered

Half the audited sites (50.7%) had at least one heading phrased as a question with no answer under it that a machine could isolate, and 61.5% had no question-and-answer content of any kind. The heading is the easy half. Answer the question in the first sentence beneath it, completely, with your business named and the conditions after; writing to be cited shows the shape of an answer block.

7. Reviews left on the platforms, or copied into the markup

64.7% of the sites showed a crawler no review signal at all, although most of those businesses had reviews on their profiles. The fix is to quote a few real ones on the pages they describe, as text, with the reviewer's first name, the date and the platform named. The mistake on the other side is copying a platform's rating into AggregateRating markup: Google's review snippet rules say not to aggregate reviews or ratings from other websites, and a business showing reviews of itself on its own site earns no review stars in any case.

8. Writing for keywords instead of answers

The one controlled study we cite, Aggarwal et al. (KDD 2024), found that keyword stuffing did not help on its benchmark, while adding citations, quotations and statistics to a source raised its visibility in generated answers by up to about 40%. A page that repeats "best plumber in town" gives an answer nothing to say. One that states the service, the area, the price range and what a customer said about the job gives it several things.

Mistakes in measuring, and in buying help

9. Measuring with one screenshot

The same question can name different businesses on different runs, so a screenshot of one answer is a sample of one. The opposite mistake is as common: rewording the questions every month, which ends the series you are trying to read. Measure with a fixed panel of buyer questions, put to ChatGPT, Claude, Gemini and Perplexity three times each, every month, as setting up prompt monitoring describes.

10. Paying for a promise

No one outside the companies that build the assistants can promise a place in their answers, so a guarantee of placement is the first thing to walk away from. Two smaller versions of the same mistake: paying for a special file sold as required, when Google's page on AI features says none is needed, and paying for FAQ markup sold for its rich result, which Google stopped showing in Search on May 7, 2026. Buy measurement and fixes, with a re-run on a date you agree in advance.

The order to fix them in

Fix in the order each step depends on the last. Access first, because a page no crawler receives hides everything else; then the business's identity, then the form the facts are in, then markup that describes the business. Answers and reviews come next, because they give a recommendation its reason. Measure before you start and every month after, so you can tell which fix moved which question.

The GEO checklist puts twenty-five jobs in that order, and the 90-day roadmap spreads them across three months. Why ChatGPT does not know your company explains the six causes behind most of these mistakes.