Eight GEO mistakes, and how to check whether you are making them

GEO here means Generative Engine Optimization: being findable and quotable by AI search — ChatGPT, Perplexity, Claude, Google’s AI results — not geography or location pages. If you want the whole picture rather than the failure modes, start with GEO for WordPress.

Every item is written the same way: what the mistake is, how to check whether you have it, and what to change. Nothing here needs a subscription or a new platform.

1. Blocking every AI crawler in one move

The mistake: someone decides "we don’t want AI taking our content" and blocks everything with an AI-sounding name. But the crawlers are not one group. Some send visitors back — OAI-SearchBot, PerplexityBot, Claude-SearchBot, Google-Extended fetch pages so they can cite and link you. Others take the content and send nothing back. Blocking the first group buys you nothing and costs you the referrals.

• Check: look at what actually visited your site over the last month, by name, before deciding anything — a crawler log is the cheapest way. Most owners are arguing about blocking AI without knowing which bots have been there at all.
• Fix: decide by behaviour, not by the word "AI", and decide it as a business question. If visitors have to come to your site to buy — to see the demo, download the file, place the order — then being mentioned is worth more than the click you lose, and allowing the training crawlers is a defensible choice. That is the call we made for this site: our robots.txt lets all of them in, on purpose. If your content *is* the product and a summary replaces the visit, the opposite is right. What is never right is deciding it by the name.

2. Treating robots.txt as a wall

The mistake: a rule goes into robots.txt and the job is considered done. A robots.txt rule is a request. The operators who publish that they honour it do; anything that chooses not to, does not, and nothing about the file stops it.

• Check: after adding the rule, look at whether that crawler still appears in your logs.
• Fix: if you need a crawler actually stopped, turn it away at the door — refuse to serve the page — and keep the robots.txt rule as the polite version. Know which of the two you are relying on.

3. Publishing an llms.txt that nothing backs up

The mistake: a file appears at /llms.txt listing pages that are thin, out of date, or that say something different from the file. The file is an index, not a substitute for content. A second version of this mistake: two plugins both want to serve /llms.txt, one wins, and the one you configured is not the one being served.

• Check: open your own /llms.txt in a browser and click through three entries. If a page does not say what the file claims, the file is the problem.
• Fix: keep the index generated from real content rather than hand-written, and confirm which plugin is actually serving it — that is a whole question of its own.

4. Leaving the answer in paragraph five

The mistake: the page opens with context, positioning, and a story, and the actual answer arrives after three screens. A human skims past it. A model summarising the page has no reason to prefer your sentence over someone else’s that came first.

• Check: read the first two sentences of the page on their own, with the heading removed. Do they answer the question the page is titled after?
• Fix: put a self-contained answer near the top — two or three sentences that make sense quoted alone, without "as mentioned above".

5. Facts that exist for humans and not for machines

The mistake: the page looks complete, and the structured data behind it is missing a field that the search engine treats as required. Nothing on screen looks wrong, so nobody notices. We did this ourselves: our product pages carried a picture all along, but the Product structured data had no image property — a required one for merchant listings. The rich result was withheld for three weeks while every page looked fine.

• Check: run your own product, article and FAQ pages through a structured data test, and read the warnings, not just the errors.
• Fix: treat the machine-readable copy of a fact as part of the page, not as decoration — if it is on screen and not in the markup, the two will drift.

6. Key facts only in an image, or only after JavaScript runs

The mistake: the price, the specification table, or the answer itself lives in a screenshot, or is fetched and rendered client-side. Some crawlers execute JavaScript; many do not, and almost none read text out of a picture. The fact is on the page for a person and absent for a machine.

• Check: view the page source (not the inspector) and search it for the number you care about. If it is not in the HTML that arrived, assume it is not there at all.
• Fix: put the fact in the HTML, and keep the image as the illustration rather than the source.

7. Numbers you cannot back

The mistake: "trusted by 10,000+ teams", "reduces workload by 70%", a review score nobody can trace. Models cross-check claims against other sources; a number that appears nowhere else lowers confidence in the whole page rather than raising it.

• Check: for each number on the page, ask where a stranger would verify it. If the answer is "they would have to take our word for it", it is working against you.
• Fix: use facts that can be checked — what is included, how many of them, what it costs, what happens next — and drop the ones that cannot.

8. A "last updated" date that is decoration

The mistake: the date on the page, or the dateModified in the markup, moves on every deploy, or never moves at all. Both make it useless. Freshness signals only work if they are true, and a page that claims to be current while describing something that changed a year ago is worse than an honestly old page.

• Check: pick three pages and ask what actually changed on the day the page says it was updated.
• Fix: move the date when the content changes, not when the site is rebuilt.

What to do first

If you only do one thing: check which crawlers have actually visited, then read your own opening two sentences as if you had never seen the page. Those two account for most of the distance between a site that gets cited and one that does not.

Rather have it done for you? We do this as a fixed-price setup — $199, three working days, and you get the same checklist to verify it yourself.