How to Avoid Common GEO Mistakes in AI Search: Content, Measurement, Vendors
Twelve common GEO mistakes across content, measurement and vendor selection, each with the check that catches it and what Canlah AI measured.
Quick answer
Canlah AI identifies twelve common GEO mistakes in this review, four each in content, measurement and vendor selection, and the four measurement mistakes are the ones that make every other number unreadable. The most expensive single error is mistake 3, publishing only for questions that already contain the brand name, because a test panel built from the same questions then reports the gap as full visibility.
Canlah AI found that error in its own audit of canlah.ai on September 7, 2026. The site was named in 39 of 39 branded answers and in 7 of 177 non-branded answers, or 4.0%, so a branded-question panel would have reported 100.0% visibility for a brand buyers could not find without typing its name.
Treat this catalogue as directional. A mistake list cannot prove that fixing any one item moves an answer, because a single run cannot establish stable share of voice or causation. Freeze a panel of buyer questions and record engine, market, language and date, then repeat it after each correction.
Disclosure: Canlah AI publishes this guide and sells the measurement service described near the end. The first-party figures come from a self-audit of canlah.ai run on September 7, 2026 and a Singapore query archive recorded the same month, plus a cited-page layout study completed on September 23, 2026. Engine operator documentation and tool pricing were reviewed on September 23, 2026 and were not verified through paid accounts.
What counts as a GEO mistake?
A GEO mistake is a repeatable process error, not a bad answer, and the twelve below split into three families: four content, four measurement and four vendor mistakes. AI answers are volatile by design, so the test is whether the same decision produces the same distortion every time it is repeated.
- Content mistakes: work published on the assumption that engines read a file, a schema block or a word count they do not use for selection.
- Measurement mistakes: a panel design that makes the reported number mean something other than the report claims.
- Vendor mistakes: buying a promise, a score or a source list no supplier can hold to across independent models.
- Not a mistake: a mention rate that moves between two runs of the same question, or absence from an engine your buyers do not use.
The twelve most common GEO mistakes and what each one costs
Each of the twelve common GEO mistakes has one check that catches it, and the table below pairs every mistake with that check and the cost of missing it.
The twelve recurring GEO mistakes, with the check that catches each one, reviewed September 23, 2026.
| # | Mistake | How to detect it | What it costs |
|---|---|---|---|
| 1 | Publishing llms.txt as a ranking lever | Ask which named engine documents reading it | Engineering time and a false sense of coverage |
| 2 | Adding schema for an AI-specific boost | Compare the markup against Google’s AI features page | Delay, while entity and source gaps stay open |
| 3 | Publishing only pages that answer branded questions | Count the pages and panel questions that contain the brand name | No non-branded reach, reported as 100.0% visibility |
| 4 | Answering each missing mention with a new blog post | Check whether the cited domain was yours or a third party | Volume, with no change in the retrieved source set |
| 5 | Running the buyer panel once | Look for a run count per question in the report | A number with no error bar, treated as a position |
| 6 | Reporting one composite score as a rank | Ask for the formula and the per-engine breakdown | Two engines averaged into a figure neither produced |
| 7 | Merging citation, mention and recommendation | Ask which of the three the headline number counts | Performance overstated on one metric, understated on another |
| 8 | Leaving the model version floating | Check whether the report names a snapshot or an alias | Instrument movement, read as market movement |
| 9 | Buying a guaranteed ChatGPT citation | Read how the guaranteed outcome would be measured | A promise over an output no supplier controls |
| 10 | Accepting a score with no raw answers | Ask for one dated prompt, answer and source list | No way to trace a fall to a cause |
| 11 | Copying a global source list into a local plan | Compare the list against your cited-domain export | Outreach aimed at domains your engines do not cite |
| 12 | Translating an English prompt set for Chinese engines | Have a native speaker read the prompts aloud | Chinese-engine numbers that measure the translation |
Source: Canlah AI probe archives dated September 7, 2026, plus engine operator documentation and public tool pages reviewed September 23, 2026. Detection methods are editorial recommendations, not independently tested procedures.
Four content mistakes, and what to publish instead
Two of the four content mistakes, llms.txt and AI-specific schema, spend budget on inputs Google’s AI features guidance says are not required, and the other two publish pages for the wrong questions or on the wrong domain.
Mistake 1: treating llms.txt as a ranking lever
llms.txt was not documented as a selection input by any named engine on the pages reviewed on September 23, 2026, so the file should not be budgeted as a ranking lever on any surface. Google’s guidance on AI features states that site owners need no new machine readable files or AI text files to appear in them, and John Mueller has said publicly that Google does not use it. Publish a plain facts page instead: Canlah AI keeps one at canlah.ai/ai-info/ so every copy of the company description starts from one wording.
Mistake 2: adding schema in the hope of an AI-specific boost
Google states there is no special schema.org structured data that a site needs to add in order to appear in AI features, so an AI-specific markup project has no documented mechanism behind it. Structured data still earns rich result eligibility and still has to match the visible text, which is where most audits find real defects. Make the markup agree with the page, fix the entity fields, then move the remaining budget to the source layer that decides retrieval.
Mistake 3: publishing only pages that answer branded questions
A site that publishes only for branded questions is found only by buyers who already know its name, and a test panel built from the same questions reports that gap as 100.0% visibility. In Canlah AI’s September 7, 2026 self-audit, 39 of 39 branded answers named the site, against 7 of 177 non-branded answers, or 4.0%.
The rendered Google leg agreed. An AI Overview appeared on 62 of 65 valid runs for Singapore and the brand was mentioned 11 times, all of them on branded questions, leaving zero mentions across 51 non-branded runs. Publish pages that answer the non-branded questions buyers ask before they know a name, and report branded and non-branded panels on separate denominators.
Mistake 4: answering every missing mention with another blog post
A missing mention is usually a retrieval problem, and one more page on your own domain does not change which domains an engine retrieves. Across 32 open Singapore buyer questions, ChatGPT, Gemini and Google AI Overviews produced 585, 908 and 389 citations, named 321 distinct domains between them and shared only 16. Read the cited-domain export first: if the answer for your category is built from third-party pages, the backlog item is presence on those pages.
Four measurement mistakes that make the numbers unusable
The four measurement mistakes change what a reported GEO number means, and all four were visible in Canlah AI’s September 7, 2026 archive: single runs, blended scores, merged metrics and floating model versions.
Mistake 5: running the buyer panel once
One run per question produces a number with no error bar, and the measured churn is large enough to swallow most reported movement. In Canlah AI’s ChatGPT archive, 7 of 20 open funnel questions returned at least one mention across three runs, and none returned a mention in all three. Set the run count before collection, state it in the report, and treat any change smaller than the run-to-run spread as unproven.
Mistake 6: reporting one composite score as a rank
A composite AI visibility score is an average of surfaces that disagree, so it is not a rank and cannot be compared across vendors with different formulas. The same 32 questions produced citation counts of 585, 908 and 389 on three engines in one day, which no single figure represents honestly. Ask for the formula, the weights and the per-engine breakdown, because a transparent lower number is more useful than an unexplained higher one.
Mistake 7: merging citation, mention and recommendation
Citation, mention and recommendation are three different events, and reporting them as one metric overstates one and understates another. Citation means the content entered the evidence pool, mention means the brand entered the answer and recommendation prominence means the brand was the pick rather than a passing reference. Record the three in separate columns from the first baseline, because retrofitting the split means re-reading every archived answer.
Mistake 8: leaving the model version floating
A floating model alias is an unrecorded variable, and it can move an answer further than a content change does. In Canlah AI’s September 7, 2026 archive, two sibling Gemini aliases answered the same question in the same minute with 5,487 and 3,903 characters, 40.6% apart, and glossed the category differently. Pin the model snapshot where the API exposes one, record the alias where it does not, and stamp every report with the version behind it.
Four vendor mistakes that appear during procurement
The four vendor mistakes appear at contract stage, when a buyer accepts a guarantee, a score, a source list or a prompt set that has not been checked against the buyer’s own engines, market and language.
Mistake 9: buying a guaranteed ChatGPT citation
No supplier controls how an independent model answers a buyer question, so a guaranteed citation is a contractual offer with conditions rather than evidence that a method works. Several Singapore providers publish citation or refund guarantees, and the clause that matters is how the outcome would be measured. Canlah AI signs clauses for a locked question pool, a stated probe volume per engine, a dated baseline, archived raw answers and a written acceptance bar (AI names the client at least 30% more often than at its starting baseline, re-tested every month on the same buyer questions), and declines to sign a guaranteed citation, AI ranking or #1 position.
Mistake 10: accepting a score with no raw answers
A score without archived answers cannot be traced to a cause, which makes every later fall an argument rather than a diagnosis. Ask any platform or agency for one dated sample containing the prompt, language, engine, complete answer, cited sources and the action created from it. Tool pricing shows what that sampling costs: Peec AI prices Starter, Pro and Advanced at 50, 150 and 350 tracked prompts and the Semrush AI Visibility Toolkit lists US$99 per month for one domain and 25 tracked prompts. AthenaHQ defines one credit as one AI response, and all three publish their sampling allowances, which is the transparency this mistake is about.
Mistake 11: copying a global source list into a local outreach plan
Global source rankings are accurate for their own samples and still wrong as a target list for a local market. The thirteen domains leading the public United States rankings reviewed on September 23, 2026 took 14 of 585 ChatGPT citations in Canlah AI’s Singapore archive, or 2.4%, and Ahrefs Brand Radar reported reddit.com at 16.8% of ChatGPT top-source citations for September 2026 while ChatGPT cited Reddit zero times on this question set.
The local picture differed by engine. ChatGPT’s most-cited domain was developers.google.com at 9.7%, Gemini’s was stridec.com at 8.5% and Google AI Overviews’ was mediaplus.com.sg at 8.5%, while local .sg domains took 26.2%, 23.6% and 29.0% of those engines’ citations.
Mistake 12: translating an English prompt set for Chinese engines
A translated prompt measures the translation, because a Chinese buyer does not phrase a comparison query as a literal rendering of the English one. DeepSeek, Qwen and Doubao also retrieve from a different source graph, built on Zhihu, WeChat public accounts, Xiaohongshu and Chinese media rather than Reddit or Wikipedia. Write the Chinese panel natively, score the output in Chinese, and keep the metric definitions identical to the English panel.
Three AI SEO myths that the evidence does not support
Three AI SEO myths survive in GEO planning in September 2026 because few teams check them against operator documentation or cited pages: that AI search has replaced SEO, that blocking Google-Extended removes a site from AI Overviews and that longer pages earn more citations.
AI search has not replaced SEO. Google AI Overviews and Google AI Mode are built on the ordinary Google Search index, so crawlability, indexing and a page that matches its markup remain the entry conditions. The unit of success changed, from a click to a citation.
Blocking Google-Extended does not keep a site out of AI Overviews. Google documents Google-Extended as a control for training and grounding in other Google systems, while the controls for Search AI features are nosnippet, data-nosnippet, max-snippet and noindex, which also cut ordinary snippets.
Longer pages do not earn more citations on the measured evidence. Canlah AI’s September 2026 layout study covered 92 pages cited by ChatGPT and Google AI Mode: they ran from 40 to 9,000 words with a median near 1,300 to 1,500, and the copied sentence was usually the first sentence of a paragraph, as on 15 of 18 legal pages.
What each crawler and preview control actually does
Six crawler and preview controls decide which AI surfaces can use a page, and OpenAI documents GPTBot and OAI-SearchBot as independent settings, so blocking GPTBot does not remove a site from ChatGPT search answers.
Crawler and preview controls, with the effect each one is documented to have, reviewed September 23, 2026.
| Control | What it does | What it does not do | Where it is documented |
|---|---|---|---|
| OAI-SearchBot | Surfaces pages in ChatGPT search answers; opted-out sites are not shown there | Does not govern training use, a separate setting | OpenAI crawlers documentation |
| GPTBot | Signals whether crawled content may train OpenAI foundation models | Does not remove a site from ChatGPT search answers | OpenAI crawlers documentation |
| PerplexityBot | Fetches pages Perplexity can cite in its answers | Does not control other vendors’ engines | Perplexity crawler documentation |
| Google-Extended | Limits AI training and grounding in some other Google systems | Does not remove a site from AI Overviews or AI Mode | Google Search Central, AI features |
| noindex, nosnippet, max-snippet | Limits what Search may show from the page, AI features included | Does not target AI features alone; ordinary snippets go too | Google Search Central, AI features |
| llms.txt | Serves as a human-readable index a site chooses to publish | Not documented on the reviewed pages as an input to any named engine | Google guidance on AI features |
“Not documented on the reviewed pages” means the operator’s public materials did not name that behaviour when reviewed on September 23, 2026. It is not proof that the behaviour is absent.
How to tell a mistake from ordinary answer volatility
Ordinary answer volatility has a measured size in Canlah AI’s September 7, 2026 archive, and anything inside it is not evidence of a mistake: 3 of 31 twice-rendered questions returned a Google AI Overview on one render and none on the repeat. Phrasing alone moved the ChatGPT mention rate for the same brand, engine and hour from 13.6% under 20 full-sentence buyer questions to 0.0% under 12 short head-term questions.
Cited sources churned further than verdicts did. Three ChatGPT runs cited 323 distinct domains, of which 36, or 11.1%, appeared in all three. Mean pairwise overlap between two runs of the same question was 0.28.
The rule follows: a mistake reproduces when the panel is repeated under recorded conditions, and volatility does not. Record the instrument as carefully as the result, because 14 of 116 ChatGPT responses were truncated at an output cap and 5 of 70 Overview renders failed in transport.
A seven-step sequence for correcting the twelve
Correct the twelve mistakes in seven steps, measurement first, because a content fix cannot be read until the panel that measures it is stable.
- Split the question pool into branded and non-branded sets and report the two denominators separately.
- Fix the run count and name the model snapshot or alias, then record engine, market, language, account state and date before collecting.
- Export the cited domains per engine and compare them against the outreach list before commissioning content.
- Separate citation, mention and recommendation into three columns and keep the raw answers behind each.
- Check the crawler and preview controls against the table above, since one robots.txt line can close one answer surface and leave the others open.
- Correct entity facts across the site, Wikidata, LinkedIn, Crunchbase and Google Business Profile before publishing new pages.
- Repeat the frozen panel after each change and describe movement as association, not cause.
What the GEO mistakes catalogue cannot prove
This catalogue of twelve mistakes cannot prove three things, and the same three limits apply to any vendor’s mistake list.
- It cannot prove that correcting any one mistake will change how an engine answers a given question.
- It cannot show that a before-and-after difference was caused by the correction, because model updates, competitor activity and index changes run in parallel.
- It cannot rank the twelve by financial cost, or confirm engine behaviour beyond what each operator documents publicly.
When a mistake is worth paying an agency to fix
Paying an agency to fix GEO mistakes is justified when the buyer panel outgrows a spreadsheet: more than one engine, more than one language, a market where an inaccurate answer creates regulatory risk or several teams publishing facts that must stay consistent. Below that threshold a documented monthly manual panel is usually enough, and the upgrade decision should follow measurement complexity, not fear of missing a trend.
Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. It is less useful below that threshold, for a team tracking one engine in one language that needs a spreadsheet panel rather than a managed service. Its own evidence is also weak: the September 7, 2026 self-audit returned an overall score of 71 and an internal usability review judged that report unfit to send to a client without repair.
Do not act on this catalogue alone. Run the checks in the first table against your own panel, ask any vendor for the raw answers behind its number and treat a mistake as confirmed only when it reproduces under recorded conditions.
Frequently asked questions
What are the most common GEO mistakes?
The most common GEO mistakes, often called AI SEO mistakes, are measurement mistakes: running each question once, reporting a composite score as a rank, merging citation with mention and leaving the model version floating. The costliest single GEO mistake is publishing only for branded questions, which a branded test panel reports as 100.0% visibility. Content mistakes such as llms.txt projects and AI-specific schema work cost less but delay the work that changes retrieval.
Are AEO mistakes different from GEO mistakes?
Mostly no. AEO and GEO describe the same buyer surfaces, so the same twelve errors apply. The difference worth keeping is scope: AEO discussions often stop at featured snippets and voice answers, while a GEO panel covers ChatGPT, Gemini, Google AI Overviews and Google AI Mode.
Can blocking GPTBot remove a brand from ChatGPT answers?
No. OpenAI documents GPTBot and OAI-SearchBot as independent settings, and states that sites opted out of OAI-SearchBot will not be shown in ChatGPT search answers. Blocking GPTBot signals only that crawled content should not train the foundation models. Check OpenAI’s crawler documentation before editing robots.txt, because this answer reflects the version reviewed on September 23, 2026.
Is Canlah AI’s own site free of these mistakes?
No. The September 7, 2026 self-audit scored canlah.ai at 71 overall, found a 4.0% non-branded mention rate across 177 answers and produced a report an internal usability review judged not ready to send to a client. The archive appears here because it shows the measurement, not because it is a strong result.
Method and sources
This catalogue was assembled from two Canlah AI probe archives dated September 7, 2026, a Canlah AI cited-page layout study dated September 23, 2026 and operator documentation and tool pricing pages reviewed on September 23, 2026. The twelve were selected because they appeared in those archives or in the reviewed vendor pages, not because a controlled study ranked them. The review did not test provider dashboards, audit client work or confirm commercial terms.
- Google Search Central, AI features and your website. No new machine readable files and no special schema; preview controls for AI features.
- OpenAI crawlers documentation. OAI-SearchBot and GPTBot as independent settings, and the effect of opting out of each.
- Search Engine Journal on Mueller and llms.txt. Public Google position on llms.txt.
- Peec AI pricing, Semrush AI Visibility Toolkit and AthenaHQ pricing. Prompt, domain and credit allowances by plan.
- Ahrefs Brand Radar. Reported top-source share for ChatGPT citations, September 2026.
- Canlah AI facts page. Measurement protocol, probe volumes and guarantee position.
- Canlah AI self-audit of canlah.ai, September 7, 2026, anonymised. Branded and non-branded mention counts, Overview renders and the overall score.
- Canlah AI Singapore query archive, September 2026, anonymised. Citation counts, cited-domain shares and run-to-run overlap.
- Canlah AI cited-page layout study, September 23, 2026: 60 queries across ChatGPT and Google AI Mode, 526 citations on 92 pages. Word-count range and copied-sentence position.
See where your brand stands in AI answers. Run the free AI visibility audit →
Related articles
FREE WHITEPAPER
Marketing in the Agent Era
All 13 chapters public — no email wall. Includes an original dataset on the agent-readiness of 50 cross-border DTC brands.
Read it free →