The five mistakes are consistent: treating GEO as keyword SEO with new tags; obsessing over ChatGPT and ignoring the other engines; mistaking brand vanity prompts for buyer prompts; outsourcing without a measurement framework; and starting before checking robots.txt. Each is fixable. Together they explain why most "GEO programmes" produce no measurable result in the first 90 days. Avoid these and the same time investment produces meaningfully different outcomes.
The pattern
We've now run hundreds of Search Visibility Audits across categories from B2B SaaS to e-commerce to professional services. The brands who hire us have usually been "doing GEO" for somewhere between zero and twelve months when they engage. The pattern of what doesn't work is depressingly consistent.
Most teams aren't lazy or incompetent. They're pattern-matching from SEO experience and finding that the patterns don't quite transfer. The result is effort being spent on the wrong work — not no work, just misallocated work.
This piece is the cheat sheet. Read it before you start, save 90 days.
The five mistakes
Treating GEO as keyword SEO with new tags
The single most common failure mode. A team is told to "do GEO," they assume it's like SEO, they identify "AI-relevant keywords," they write blog posts targeting those keywords, and they wait for AI engines to start citing them.
Months later: nothing. The blog posts rank okay on Google but never appear in AI engine answers. The team is confused. They blame the AI engines for being unpredictable.
What's actually happening: GEO doesn't have keywords in the SEO sense. AI engines aren't matching query strings to page strings — they're synthesising answers from a corpus of authoritative sources about your category. Your own website is one source among hundreds. Posting more content on your own site is a fractional improvement; what would actually move the needle is being mentioned in the dozens of other sources the AI is reading.
The fix: stop optimising your own content as the primary GEO motion. Start asking "what would the AI need to encounter, where, before it confidently mentions us?" The answer almost always involves Digital PR, editorial relationships, Wikipedia work, and third-party content you don't own.
Counter-pattern: If 80% of your GEO budget is going into your own site's content, you're doing SEO and calling it GEO.
Fix horizon: 1 quarter to redirectObsessing over ChatGPT, ignoring the other engines
The CMO uses ChatGPT. The team uses ChatGPT. The team starts measuring AI visibility — in ChatGPT. They optimise for ChatGPT. They report ChatGPT numbers to leadership. The other engines barely come up.
Two problems with this. First, your buyers don't all use ChatGPT. Enterprise B2B audiences over-index on Claude. Research-heavy buyers (analysts, journalists, technical evaluators) over-index on Perplexity. Consumer audiences increasingly encounter AI Overviews as the default. Optimising only for the engine you personally use is optimising for one channel and missing four.
Second, the engines work differently enough that ChatGPT-specific optimisation doesn't transfer cleanly. ChatGPT leans heavily on training data. Perplexity is almost entirely live retrieval. Gemini integrates Google Search results. AI Overviews has its own logic. A strategy that works for ChatGPT may produce zero result on Perplexity, and vice versa.
The fix: your prompt set runs across ChatGPT, Claude, Perplexity, Gemini, and AI Overviews. You report SoAIV by engine, not just an aggregate. You weight your effort by which engines your buyers actually use, not which engines your CMO uses.
Fix horizon: 1 day to start measuring properlyMistaking brand-vanity prompts for buyer prompts
This one is subtle. A brand wants to test their AI visibility. They ask ChatGPT: "Tell me about [their brand name]." ChatGPT obligingly gives them a paragraph. They conclude they're visible.
They're not visible. They're findable. Those are different things.
Asking the AI directly about your brand by name is the equivalent of Googling your own company name and looking at the top result. Of course you appear — the query is literally about you. What matters is whether you appear in queries about your category, when buyers haven't decided yet what to consider.
Real buyer queries look like:
- "What's the best [category] for [use case]?"
- "Compare [competitor A] and [competitor B]"
- "What should I look for when choosing a [category]?"
- "Are there alternatives to [established player]?"
Your brand name doesn't appear in any of them. The question is whether the AI's answer to those queries includes your brand. That's the test that matters.
The fix: build your prompt set from real buyer language, not from your brand. If your prompt set includes your own brand name, you're testing findability, not visibility. Strip them out.
Fix horizon: A few hours to rebuild the prompt setOutsourcing without a measurement framework
An agency pitches GEO services. The brand signs a 6- or 12-month contract. The agency starts producing "AI-optimised content," "AI visibility audits," and monthly reports full of activity metrics. Six months later, the brand can't tell whether anything has actually changed in their AI visibility.
The structural problem: the agency is the one producing the measurement, against criteria they themselves set. The reports show activity (content produced, mentions earned, audits run) without an independent measure of outcome (Share of AI Voice change). When you ask "but did our SoAIV actually move?" the answer is rarely a clean number.
The fix: maintain your own outcome measurement, separate from your agency's deliverables. Run a quarterly Share of AI Voice baseline yourself (or with an independent third party). Hold the agency accountable to that number, not to their own activity reports.
This isn't about distrust of agencies. The good ones welcome external measurement because it makes their work look better when they're succeeding. The agencies that resist external measurement are the ones who know their work isn't producing results.
Fix horizon: 1 quarter to set up parallel measurementStarting before checking robots.txt
Easiest mistake to fix; surprisingly common. A brand spends months on Digital PR, content production, schema markup, Wikipedia work — all to influence AI visibility. Meanwhile, their robots.txt file (or hosting platform's defaults) is blocking the AI crawlers that would otherwise index their content for live retrieval.
The result: the live-retrieval layer is closed off. All the work is going into the training-data layer (which updates slowly) while the fast-feedback layer is structurally inaccessible.
The crawlers most commonly blocked accidentally:
GPTBot— OpenAI's crawlerClaudeBot— Anthropic's crawlerPerplexityBot— Perplexity's crawler (the one most often blocked, because some hosts include it in default lists)Google-Extended— Google's AI training and AI Overviews retrieval crawlerCCBot— Common Crawl bot, which underpins many AI training datasets
The fix: take 5 minutes. Open yourdomain.com/robots.txt. Search for any of the user-agent strings above. If any are blocked, decide whether the block was intentional. If not, remove it. Cheapest possible GEO improvement, and the only one in this list that pays off in days rather than months.
(There are legitimate reasons to block specific crawlers — IP protection for paid content, reducing crawl load on small sites, etc. But "we accidentally inherited a config" is a poor reason and easily fixed.)
Fix horizon: 5 minutesWhy these mistakes happen
Each of the five mistakes is rooted in a specific structural pressure on marketing teams. Understanding the pressure helps you avoid the mistake even after the obvious symptoms fade.
Mistake 1 happens because SEO instincts run deep.
Most marketing teams have spent years optimising their own content. The muscle memory is "produce the page, optimise the page, watch the page rank." GEO requires a different instinct — "what about the third-party sources?" — and it's slower to develop. The fix is partly intellectual (understanding why) and partly organisational (giving Digital PR people seat at the GEO table, not just SEO people).
Mistake 2 happens because of personal usage bias.
If you and your CEO use ChatGPT, you'll over-weight ChatGPT in your strategy. The bias is honest but expensive. The discipline is to ask "what do buyers use?" — which is a research question, not an introspective one. Survey your customers about which AI tools they use. The answers usually surprise.
Mistake 3 happens because brand-vanity prompts feel reassuring.
It's psychologically comforting to test prompts that include your brand name and see your brand mentioned. The relief is real. The relief is also illusory — those prompts don't reflect buyer behaviour. The discipline of testing with real buyer language exposes weakness, which is exactly why people avoid it.
Mistake 4 happens because external measurement is uncomfortable.
Holding agencies accountable to outcome measurement they don't control is harder than reading their reports. It requires the brand to do work, to hold tension in the relationship, to potentially fire an underperforming agency. Most brands prefer the comfortable fiction of activity reports.
Mistake 5 happens because robots.txt isn't sexy.
Nobody's career was made by checking robots.txt. The mistake persists because the work that would catch it is unglamorous and falls between marketing and engineering. Both teams assume the other has handled it. Neither has.
The 90-day plan that actually works
If you're starting GEO work from scratch with 90 days to show progress, the work that actually produces measurable results:
The sequence we'd run
Days 1–7: Establish baseline. Run a 50-prompt SoAIV measurement across ChatGPT, Claude, Perplexity, Gemini, AI Overviews. Tag by funnel stage. Parse mentions. This is your starting line.
Days 1–7 (parallel): Run the technical audit. robots.txt, AI crawler access, schema markup, JS-rendering check, comparison-content audit. Fix anything broken. These are the only quick wins available.
Days 8–30: Identify the gap. From the baseline, pick the one biggest weakness — usually one of: missing Wikipedia, weak editorial volume, sentiment problem, missing decision-stage content. Build a focused 60-day plan against that single weakness.
Days 31–60: Execute the focused plan. Don't try to fix everything. Pick the highest-leverage gap and pour effort into it. Sustained effort on one area outperforms scattered effort across five.
Days 61–90: Continue executing + early measurement. Some changes will start showing up in live retrieval. Training-data changes won't appear yet — that takes longer. Set expectations accordingly.
Day 90: Re-measure. Compare to baseline. Where did movement happen? Where didn't it? Use the data to decide what to keep, what to change.
- Don't expect dramatic SoAIV growth in 90 days — the systemic changes take 6–12 months
- Do expect technical-fix wins to show up fast — within weeks, often
- Do expect early signal on whether your hypothesis was right — even if absolute numbers haven't moved much
What success looks like at the 90-day mark
Calibrating expectations matters because GEO sits in an awkward zone — fast enough to measure quarterly, slow enough that you can't expect 6× growth in 90 days. Realistic outcomes after a focused first 90 days:
- Technical fixes deployed. robots.txt, schema, comparison content. These are infrastructure improvements that pay off later.
- Live-retrieval citations starting to appear. If you've published strong AI-citable content, you'll see it cited in Perplexity within weeks, in ChatGPT browse mode within a month or two, in Claude's web tool similarly.
- Editorial pipeline established. Maybe 8–15 high-authority mentions earned in the period. The flywheel hasn't fully spun up yet, but the motion exists.
- Wikipedia work in progress. Probably not a published article yet — that takes longer — but a draft, a notability case being built, or improvements to an existing article underway.
- Baseline + one re-measurement. You can compare quarter 1 to quarter 0. The comparison is your reality check.
What you won't see at 90 days: dramatic Share of AI Voice growth in ChatGPT or Claude. Those changes mostly require training-data updates, which happen on 6–12 month cycles. Expecting otherwise leads to false disappointment and bad decisions.
The mistake we haven't made the list
Honourable mention for the mistake that's increasingly common but not quite top-five frequency yet: relying on AI-generated content to "do GEO."
Some teams, when told they need more content, default to generating it with AI. The result is content that's grammatically fine, factually plausible, and instantly recognisable as AI-generated to both human readers and AI engines themselves. AI engines specifically downweight content that pattern-matches to AI-generation. So you've spent the budget, produced the content, and made yourself slightly less visible by polluting your own domain with low-trust content.
If we revisit this list in 12 months, this will probably be in the top 5. Today it's a watchlist item.
Closing thought
GEO isn't conceptually difficult. The mistakes above are obvious in retrospect. The reason they're so common is that the discipline is new enough that nobody's fully calibrated to the differences yet — most teams are running SEO playbooks with new vocabulary.
The teams that get this right early have a window. AI search is in a phase where being deliberate is unusually rewarded; the gap between "we're trying" and "we have a plan" is wider than it is in mature channels. The brands that close that gap in 2026 will compound on it through 2027.
Most won't. That's the opportunity.