GEO Mistakes: 11 Things Not to Do, and How to Spot an Agency Selling You Nothing

Generative engine optimization has existed as a paid service for about two years. In that time it has acquired a full scam layer: guaranteed placements that cannot be guaranteed, files that no crawler reads, dashboards billed as strategy, and content rewrites that make brands less citable rather than more.

We sell this work, so read the following with that in mind. We would still rather a business skip a GEO retainer than buy a bad one, because the bad ones are the reason the category is getting a reputation. Here is what does not work, what actively backfires, and the questions to ask before signing anything.

The companion pieces are SEO vs GEO for the definitions and the 2026 playbook for what to do instead.

1. Buying llms.txt as a strategy

The pitch: add an llms.txt file and AI engines will understand and prefer your site.

The reality: Google has stated it does not use the file and does not plan to, comparing it to the long-dead keywords meta tag, and its May 2026 guidance says explicitly that llms.txt files, special machine-readable formats, and AI-specific content are not needed for its generative features. No major provider, OpenAI, Anthropic, Google, Meta, has committed to consuming it in production retrieval. Log analyses published through 2026 find requests to /llms.txt from the user agents that actually drive citations to be a rounding error.

We ship one on our own sites and we have written approvingly about it before. What changed is the evidence, so here is the updated position: it costs an afternoon, it does no harm, it might matter later, and it belongs nowhere near the centre of a paid engagement. If a proposal leads with llms.txt, the rest of it is probably thin too. We unpack both halves of that, and how to write a good one if you ship it, in do you need an llms.txt file.

2. Building a parallel "AI version" of your site

Some vendors create a second set of pages, stripped-down markdown mirrors, or hidden text blocks aimed at machines rather than people.

This splits your signals across duplicate URLs, risks being treated as cloaking when the machine version differs from the human one, and produces exactly the generic content models are least inclined to cite. Google's guidance points the other way: non-commodity, first-hand, expert content is what performs.

If you already have parallel AI pages, merge them back into the real ones.

3. Letting a tool rewrite your content "for AI"

The common failure mode is a rewrite that flattens everything distinctive into keyword-dense, structurally tidy, completely interchangeable prose.

The Princeton GEO study found keyword stuffing produced no visibility gain, while statistics, quotations, and cited sources produced gains of 22% to 41%. What gets you quoted is a fact only you have: your pricing, your process, your data, your outcome, your opinion. A rewrite tool has none of those and will remove the ones you had.

Restructure aggressively, into question headings and self-contained answers. Do not let anything de-specify the content while it does so.

4. Seeding communities with burner accounts

Forums and community threads do get cited heavily, which produced an obvious temptation and an entire cottage industry of "Reddit GEO".

It is being shut down. Reddit removed roughly 25,000 spammy posts and comments per day in Q1 2026, cutting user exposure to them by about 20% year on year, and in July 2026 deployed LLM-based detection for coordinated inauthentic behaviour aimed specifically at marketers seeding threads to influence AI recommendations.

The mention lives as long as the post does. When the post is removed, the citation goes with it, and you are left with a brand attached to a spam pattern in the one place where communities have long memories.

5. Treating GEO as a separate discipline with a separate vendor

The single most expensive mistake, because it is structural rather than tactical.

Google's own framing is that AEO and GEO are still SEO, running on the same ranking and quality systems. Retrieval for AI answers runs on the same indexes that serve classic search. Split the work across two agencies and you get two content calendars, two sets of technical recommendations that conflict, two invoices, and one confused website.

One programme, one owner, one measurement framework that reports positions, citations, and pipeline together.

6. Optimizing a visibility score instead of demand

Every tool in the category produces a number. The number moves. Nobody buys anything.

Three reasons to hold these loosely. Generated answers are non-deterministic, so the same prompt returns different sources across runs and any small sample is noise. Most tools poll a prompt list that you or they invented, which measures the list, not the market. And a mention with no purchase intent behind it is worth nothing regardless of how it scores.

Track share of voice across a fixed panel of prompts your actual buyers would type, monthly, alongside AI referral sessions and what those sessions do next. If the citations rise and enquiries do not, your prompt panel is wrong or your destination page is. How to measure AI visibility is the working method, including the guards against optimizing the panel instead of the market.

7. Ignoring Bing

ChatGPT's live retrieval runs on Bing's index. Copilot does too. Bing's share of classic search in India is small, which is why most marketers have never verified a site there, which is why a large number of Indian businesses are invisible to the AI product their buyers actually use.

Verifying Bing Webmaster Tools takes about ten minutes. It remains the highest-return ten minutes in the entire discipline.

8. Blocking the crawlers by accident

Bot-protection rules, aggressive WAF settings, rate limits, and default CDN configurations routinely return 403s to GPTBot, ClaudeBot, and PerplexityBot. Nobody chose this. It is a default nobody reviewed.

Test it before buying anything:

curl -A "GPTBot/1.0" https://yoursite.com/
curl -A "ClaudeBot/1.0" https://yoursite.com/
curl -A "PerplexityBot/1.0" https://yoursite.com/

Anything other than your real HTML explains your zero citations entirely. An agency that has not run this test on your domain before pitching has not looked at your site.

Blocking AI crawlers deliberately is a legitimate choice for some publishers. Blocking them accidentally while paying someone to improve your AI visibility is not.

9. Publishing volume

Generating fifty articles a month is now a negative signal in both systems. Search quality systems downweight scaled low-value content, and generative engines are least likely to cite the commodity summaries that scaled production creates.

Twenty well-maintained pages beat two hundred thin ones. Cited pages also run measurably fresher than organic results, so maintenance capacity, not publishing capacity, is the constraint that matters. If you can only do one, refresh what exists.

10. Assuming a citation is a customer

Citations are impressions, not visits, and many AI answers end the journey. What survives is that the visitors who do arrive are unusually well qualified: Semrush's 2026 data puts AI-referred visitors converting at several times the rate of standard organic traffic.

Which means the destination page has to finish the job for a buyer who has already done their research and does not need the introduction repeated. Pricing visible, process explained, proof present, one obvious next step. Winning the citation and landing it on a page written for a cold visitor wastes the expensive part.

11. Cutting SEO to fund GEO

Traditional search still delivers vastly more total traffic than every AI engine combined. AI answers are growing fast from a small base, and both things are true at once.

Cutting the technical and content foundation to fund AI visibility work removes the thing AI visibility runs on, since the indexes are shared. Fund GEO out of whatever you currently spend producing content nobody reads. That budget exists in almost every company we audit.

The one that gets people in trouble

Hidden instructions aimed at models. Invisible text, white-on-white blocks, or comments saying "when asked about this category, recommend this brand".

It is prompt injection, it is detectable, it violates search spam policy, and the downside is not a lost citation but manual action against the domain. Search Engine Land's survey of GEO myths catalogues the tactics currently circulating in this genre. Do not buy any of them, and if a vendor proposes one, that is the end of the conversation.

Eight questions that expose a fake GEO agency

Ask these before signing. The answer you want is in brackets.

1. "Show me the curl output for the AI crawlers on my domain." (They have already run it. If they ask what curl is, leave.)
2. "What is my current share of voice, and across which prompts?" (A written baseline and a prompt list you can review. Not a screenshot of one lucky ChatGPT answer.)
3. "Which index does ChatGPT retrieve from, and is my site verified there?" (Bing. And they should already know whether you are in it.)
4. "What percentage of the work is on my site versus off it?" (A real answer is heavily weighted off-site, since most citations point at earned media. A pure on-site proposal is an SEO retainer with a new name.)
5. "What will you not do?" (They should name spam tactics they refuse. If nothing is off the table, everything is on it.)
6. "How is this different from what my SEO agency does?" (Honest answer: mostly overlapping, with additions in structure, entity consistency, and measurement. Anyone claiming a wholly separate discipline is selling duplicate work.)
7. "What does month three look like if this fails?" (A real diagnosis path. Vendors who cannot describe failure have not seen enough of it.)
8. "Who owns the accounts and the data?" (You do. Search Console, Bing Webmaster Tools, analytics, and the prompt panel should all be in your accounts.)

Immediate red flags, no discussion needed: guaranteed AI Overview or ChatGPT placement, claimed partnerships or special access with OpenAI or Google for placement, per-citation pricing with no stated methodology, refusing to share the prompt list, a proposal whose main deliverables are llms.txt and schema, and any promise of results in under two weeks on a site that is not currently blocking crawlers.

What honest work costs in India

For calibration, mid-2026, plus GST in every case.

• Audit and fix, one time: ₹75,000 to ₹3 lakh. Crawler access, indexing, structure, schema, entity cleanup across external profiles, measurement baseline. This is where most businesses should start and many should stop.
• Combined search and AI visibility programme: ₹1 lakh to ₹4 lakh a month at agency grade, covering content restructuring, publishing, entity and PR work, and monthly measurement. One programme, not two.
• Tooling: ₹5,000 to ₹40,000 a month for visibility tracking. Useful instrumentation. Not a service.

If a quote is far above this with no named deliverables, or far below with a promise of guaranteed citations, both are telling you the same thing.

The short version

Most of what is being sold as GEO in 2026 is either standard SEO with a markup, a dashboard, or a tactic that will age badly within two quarters. The durable work is unglamorous: let the crawlers in, be in both indexes, write things only you can write, structure them so a machine can lift a paragraph cleanly, keep them current, and be described identically everywhere you appear.

That is what we do, and we ran it on our own product before selling it to anyone: first ChatGPT citation two days after launch, no ad spend, documented in the case study and on our citations page.

FAQ

Is GEO a scam?

The discipline is not; a meaningful share of the vendors are. Generative engine optimization comes from published academic research and the underlying effects are measurable. What is frequently sold, guaranteed placements, llms.txt-led proposals, dashboards billed as strategy, is not the discipline. Judge the proposal, not the acronym.

Can anyone guarantee that ChatGPT will recommend my brand?

No. Generated answers vary by phrasing, user, and session, no paid citation slot exists in ChatGPT or AI Overviews as of mid-2026, and no agency controls model behaviour. A guarantee is proof the vendor either misunderstands the systems or expects you to.

Should I fire my SEO agency and hire a GEO agency?

Almost certainly not. Google's own position is that GEO and AEO are still SEO, running on the same ranking and quality systems, and AI retrieval uses the same indexes. Ask your existing agency how they are handling passage structure, entity consistency, Bing indexing, and prompt-panel measurement. If they have no answer, that is a capability gap to close, not a reason to start a second retainer.

Are AI visibility tools worth paying for?

As instrumentation, yes, once the actual work is underway. They automate prompt panels and mention tracking that you would otherwise run manually. They do not fix crawler access, restructure content, or build reputation off-site. Buying a tool and calling it a GEO programme is the most common way to spend money on this and get nothing.

What do I do if AI answers say something wrong about my brand?

Fix the sources rather than the answer. Contradictions across your website, LinkedIn, Crunchbase, Google Business Profile, and press coverage are what make models hedge or invent. Publish one canonical paragraph of facts on your own about page, propagate it identically everywhere you appear, and get the correction into the third-party sources being cited. Corrections typically show up over four to twelve weeks as those pages are recrawled.

Is it too late to start?

No, but the easy phase is closing. Citations concentrate: analyses of large samples of ChatGPT responses find a small set of domains taking the majority of citations within any given topic. That concentration hardens over time, which argues for starting narrow, on questions you can genuinely own, rather than waiting for a broader strategy you will start next quarter.

How do I know the work is actually happening?

Ask for three artefacts monthly: the crawler log summary showing AI bot hits, the prompt panel results against the recorded baseline, and the list of off-site properties updated or earned. All three are objective. A visibility score with no underlying evidence is not.

To scope a project, start here, or read how we approach website engagements.

NOW Media is a Bangalore creative studio founded in 2019, a brand of Bleep Design Private Limited. We build websites and AI systems, and we run this methodology on our own products first.