The GEO tactics most agencies miss (and the ones they oversell)

Ask most agencies for a GEO audit and you’ll get the same pack back: an llms.txt file and a promise to “chunk” your content so AI can read it. Google’s own guide to optimising for generative AI search, updated in July 2026, lists both under things you can ignore for Google Search. Oh dear.

The GEO tactics that actually change whether ChatGPT, Gemini, Copilot and Google’s AI Overviews cite you are less glamorous. They sit in CDN settings, server logs, a Search Console toggle most people haven’t opened, and what strangers write about you on Medium. This guide covers the eleven we see missed most often, with the evidence behind each one, and finishes with the tactics you can stop paying for.

On this page

• Why most GEO advice starts in the wrong place
• 1. Checking robots.txt but not the CDN
• 2. Treating every AI bot as the same bot
• 3. Never opening Search Console’s generative AI control
• 4. Auditing the rendered page instead of the raw HTML
• 5. Ignoring the 404 problem (twice over)
• 6. Writing for the prompt instead of the fan-out
• 7. Selling the Princeton paper as gospel
• 8. Leaving the brand story to strangers
• 9. Chasing links when mentions matter more
• 10. Forgetting the feeds
• 11. Reporting AI “rankings” that don’t exist
• What you can stop paying for
• FAQ

Why most GEO advice starts in the wrong place

Generative engine optimisation (GEO) is the work of getting your brand retrieved, cited and described accurately in AI-generated answers. The problem with most GEO advice is that it starts at the last step, rewriting the page, and skips the steps that decide whether the page is ever considered.

Every AI answer engine works in roughly the same order. It works out what the question really is, runs searches, retrieves a handful of pages, then writes an answer from them. Google says plainly that AI Overviews and AI Mode rely on retrieval from its core search ranking systems, and that a page must be indexed and eligible for a snippet to appear at all. If a crawler can’t reach you, can’t read you or never retrieves you, no amount of clever phrasing helps.

Think of it like a job interview. Polishing your answers is worthwhile, but not if your CV never reached the hiring manager. Most of what follows is about getting the CV through the door.

1. Checking robots.txt but not the CDN

Most GEO audits confirm that robots.txt allows AI crawlers and stop there. That misses the most common blocker we find: the content delivery network or hosting firewall sitting in front of the site.

Google’s own checklist for AI features asks you to make sure crawling is allowed in robots.txt and by any CDN or hosting infrastructure. The second half matters because, since July 2025, Cloudflare blocks AI bots by default on the sites it serves, while letting customers allow or refuse crawlers stage by stage (training, fine-tuning and inference). A site can have a perfectly welcoming robots.txt and still turn every AI search bot away at the door.

What to do: check your CDN’s bot settings, your WAF rules and any “bot fight” features, not just robots.txt. Then confirm in your server logs that AI search crawlers are actually receiving 200 responses. If you’re on Cloudflare and want AI visibility, make an active choice about which crawlers to allow rather than living with whatever the default was on the day the site was set up.

2. Treating every AI bot as the same bot

Agencies often talk about “the AI crawler” as if there were one. There are several per company, and they do different jobs. Blocking the wrong one can quietly remove you from AI search while doing nothing to stop model training.

OpenAI is the clearest example. Its crawler documentation separates three user agents:

• OAI-SearchBot surfaces sites in ChatGPT search. OpenAI says sites that opt out of it won’t be shown in ChatGPT search answers, though they can still appear as navigational links.
• GPTBot crawls content that may be used to train OpenAI’s foundation models. Disallowing it is a training opt-out, not a search opt-out.
• ChatGPT-User fetches pages when a person asks ChatGPT to, and robots.txt rules may not apply because the visit is user-initiated.

OpenAI also notes that robots.txt changes for search take around 24 hours to apply. Google works the same way in principle: Googlebot governs AI Overviews and AI Mode, while Google-Extended controls training and grounding in some of Google’s other AI systems.

What to do: write robots.txt rules per user agent with a clear purpose for each. A sensible default for most commercial brands is to allow search and retrieval bots, then make a deliberate business decision about training bots. Write that decision down so the next developer doesn’t “tidy up” your robots.txt and delete your ChatGPT visibility with it.

3. Never opening Search Console’s generative AI control

Since 31 August 2026, every site in Google Search Console has a Search generative AI control under Settings. It decides whether your links and content can appear in AI Overviews, AI Mode and generative features in Discover. In our experience, very few agencies have looked at it.

The default is “include”, so most sites are fine. The risk sits in the inheritance rules. A property inherits the setting of its closest parent that has been configured, so a subfolder or subdomain property can be excluded because someone changed a setting higher up, or a child property can be overridden by a well-meaning team in another office. Google says an excluded site receives no traffic or impressions from these features, and that exclusion takes effect within days.

There’s a useful diagnostic here too. If the new Generative AI performance report shows nothing for your site, Google lists exclusion as one possible reason.

What to do: check the control on every property you manage, including URL-prefix and subdomain properties. Note that the control doesn’t affect AI training, so it isn’t a substitute for Google-Extended if that’s what a client is worried about.

Want to know if AI can see your site? Our GEO team will run your market’s prompts across the major models and send you the raw output, including who they name instead of you.

4. Auditing the rendered page instead of the raw HTML

A page can rank well in Google and still look blank to ChatGPT. Google renders JavaScript; most AI crawlers don’t.

The most detailed public data we’ve found comes from Vercel and MERJ, who monitored crawler traffic across Vercel’s network. They found none of the major AI crawlers rendered JavaScript, including OpenAI’s three bots, ClaudeBot, PerplexityBot, Meta’s crawler and ByteDance’s. GPTBot and Claude did download JavaScript files, but never ran them. Gemini, which uses Googlebot’s infrastructure, could render pages fully. That research dates from December 2024, and crawlers change, so treat it as a strong warning rather than a permanent law.

The practical consequence is nasty. Product specs and pricing tables that load after the page renders can be invisible to the crawlers behind ChatGPT, Claude and Perplexity, even while Google sees them perfectly.

What to do: stop auditing in a browser. Fetch the raw HTML with JavaScript switched off, or with curl, and check that your main content, headings, prices and FAQs are all there. If they aren’t, move critical content to server-side rendering or static generation and keep client-side scripts for extras like chat widgets. For Shopify and headless builds, this is a conversation for your development team, not your content writers.

5. Ignoring the 404 problem (twice over)

AI systems hit broken URLs far more often than Google does, from two directions: their crawlers request dead pages, and their answers send real people to pages that never existed.

On the crawler side, Vercel found ChatGPT spent 34.82% of its fetches on 404 pages and Claude 34.16%, against 8.22% for Googlebot. ChatGPT spent a further 14.36% of its fetches following redirects. Every wasted fetch is one that didn’t land on a page you wanted read.

On the answer side, Ahrefs checked 16 million URLs cited by AI assistants and found they send visitors to 404 pages 2.87 times more often than Google Search. ChatGPT was the worst offender, with 2.38% of its cited URLs returning a 404. Many of these are plausible-looking URLs that fit a site’s pattern but never existed.

Ahrefs is honest that this is a small lever for most sites, and so are we. It’s cheap to fix, though. Say 300 people a month arrive at a hallucinated URL and your enquiry page converts at 3%: that’s nine potential enquiries a month walking into a dead end (an illustrative scenario, not a benchmark).

What to do: keep redirects and XML sitemaps tidy so crawlers stop wasting requests. Filter your analytics for AI referrers, find landing pages that return 404s, and 301 redirect the ones with real traffic to the closest genuine page. For the rest, make your 404 page useful, with links to your most important services.

6. Writing for the prompt instead of the fan-out

Most GEO content plans target the question a person types. AI search engines answer a different set of questions: the sub-queries they generate behind the scenes.

Google calls this query fan-out. Its own example: a search for “how to fix a lawn that’s full of weeds” might fan out into “best herbicides for lawns”, “remove weeds without chemicals” and “how to prevent weeds in lawn”. Those sub-searches pull in pages that don’t rank for the original question at all. That’s why a site can be invisible for the head term and still get cited in the answer.

Here’s the trap. The lazy response is to spin up a separate page for every possible fan-out query. Google explicitly warns that doing this mainly to manipulate rankings or AI responses breaks its scaled content abuse policy. It’s also a waste of money.

What to do: map the likely sub-questions for each priority topic, then answer them properly inside your strongest pages, each in a clearly headed section that makes sense on its own. Use Bing’s grounding query data (more on that in point 11) to see the phrases AI systems actually used to find you.

7. Selling the Princeton paper as gospel

Almost every GEO pitch deck cites the same 2024 study. It’s a good study. It just doesn’t prove what most agencies use it to prove.

The original GEO paper from Princeton and IIT Delhi researchers, published at KDD 2024, found that adding quotations, statistics and cited sources could boost a page’s visibility in generative answers by up to 40%. Keyword stuffing, the old SEO trick, scored below doing nothing at all. That last finding alone is worth putting in front of anyone still selling keyword density.

Two details usually get left out. First, the researchers took the top five Google results for each query and rewrote one of them. The gains apply to a page that has already been retrieved. Second, when every source used the same methods, the top-ranked page’s visibility fell by 30.3% on average with the Cite Sources method, while the fifth-placed page gained 115.1%. Rewriting reshuffles a fixed pool of sources rather than getting you into it.

Then came the replication. C-SEO Bench, presented at NeurIPS 2025, tested these methods across more tasks and domains and with many sites competing at once. The authors found most current conversational SEO methods were largely ineffective and frequently hurt rankings, while traditional SEO strategies that improve a page’s position in the AI’s retrieved context were significantly more effective. Gains also shrank as more competitors adopted the same tricks.

What to do: keep adding real statistics, named sources and expert quotes, because they make content better for people and give AI systems something specific to cite. Just don’t sell it as the whole job. Retrieval, which is mostly SEO and authority, comes first. That’s why we run GEO, AEO and traditional SEO as one programme rather than three.

8. Leaving the brand story to strangers

AI assistants will describe your brand whether you’ve told them anything or not. If you haven’t published a clear, specific official version, they’ll take whatever detailed story they can find.

Ahrefs tested this by building a fake luxury paperweight brand and running 56 questions past eight AI tools over two months. It then published an official FAQ alongside three conflicting fake stories, including a Medium “investigation”. The result: Gemini, Grok, AI Mode, Perplexity and Copilot all trusted the Medium article over the official FAQ. ChatGPT held up best, citing the FAQ in 84% of its answers. The pattern was consistent: when the choice was between a vague truth (“we don’t share production numbers”) and a specific fiction (“634 units in 2023”), most models picked the fiction.

It’s a test with one fake brand, so don’t over-read the percentages. The direction is hard to argue with, though.

What to do:

• Publish a specific, dated facts page: founding year, locations, team size, what you do and don’t offer, pricing ranges where you can share them. Ranges beat silence.
• Answer the rumours directly. “We have never been acquired” is quotable; “we’re a trusted partner” is noise.
• Set alerts for your brand name alongside words like “investigation”, “former employee”, “lawsuit” and “review”.
• Ask each major assistant “What do you know about [your brand]?” every month, and log the answers.

This is the heart of what GEO actually means at Castle: being cited is only half the job if the sentence you’re cited in is wrong.

9. Chasing links when mentions matter more

Link building still matters for SEO, and SEO still feeds retrieval. But if you’re budgeting purely for AI visibility, the evidence points towards being talked about, linked or not.

Ahrefs studied 75,000 brands and found that branded web mentions correlated far more strongly with AI Overview visibility (0.664) than backlinks did (0.218). The top three factors were all off-site: mentions, branded anchor text and branded search volume. Paid search spend barely registered. The gap at the top was stark too. Brands in the top quarter for web mentions averaged 169 AI Overview mentions, against 14 for the next quarter down, and 26% of brands had none at all.

Correlation isn’t causation, and Ahrefs says so. Brands that get talked about tend to be bigger and better known for lots of reasons. But it’s consistent with how language models work: they learn what a brand is from the words that keep appearing around it.

Before anyone reaches for a mention farm, Google has a warning: seeking inauthentic mentions isn’t as helpful as it might seem, because the same systems that rank content also block spam.

What to do: put real budget into digital PR and original data people want to quote. Reviews, industry directories and genuine community presence all count too. Measure branded mentions as a KPI alongside links.

Struggling to get mentioned? See how we’ve done it for other brands in our case studies, or talk to the team.

10. Forgetting the feeds

For ecommerce and local businesses, some of the most valuable AI visibility doesn’t come from web pages at all. It comes from structured feeds and profiles that agencies often leave to someone else.

Google says AI responses can include product listings and local business information, and that Merchant Center feeds and Google Business Profiles help products and services appear in them. Microsoft makes the same point for Copilot, recommending Bing Places for Business so addresses, hours and contact details stay current in AI answers. Most UK businesses have a Google profile; far fewer have touched Bing Places.

ChatGPT is building its own route. OpenAI’s agentic commerce programme starts with a structured product feed covering titles, descriptions, images, price and availability, and recommends uploading a full snapshot at least daily. It’s currently open to approved partners, who apply through a form.

What to do: audit Merchant Center and Business Profile data with the same rigour as the website, because AI answers will repeat a wrong price or opening time with total confidence. Claim and complete Bing Places. If you sell online, apply for ChatGPT’s merchant programme and get your feed in order before the queue gets longer.

11. Reporting AI “rankings” that don’t exist

We still see GEO reports from other agencies with a neat “position 3 in ChatGPT” figure on the front page. It doesn’t mean much, because AI answers change almost every time you ask.

SparkToro and Gumshoe had 600 volunteers run 12 identical prompts through ChatGPT, Claude and Google’s AI nearly 3,000 times. The odds of getting the same list of brands twice were under 1 in 100, and the same list in the same order closer to 1 in 1,000. What did hold up was visibility percentage: some brands appeared in 60% to 90% of responses for a given intent even as their position jumped about.

Meanwhile, two free first-party data sources have arrived, and many agencies still aren’t using them:

• Google Search Console’s Generative AI performance report shows impressions from AI Overviews and AI Mode by page, country and device. It also has a “Web: multimodal” filter for searches made with Google Lens or Circle to Search, a channel almost nobody optimises images for yet.
• Bing Webmaster Tools’ AI Performance dashboard, launched in February 2026, shows total citations, cited pages and grounding queries across Copilot, Bing’s AI summaries and partner integrations. Grounding queries are the phrases the AI used to retrieve your content, which is about as close to fan-out data as anyone publishes. Microsoft also recommends IndexNow so AI answers pick up your latest content.

What to do: report share of visibility across a fixed prompt set run many times, plus first-party impressions and citations, plus what the models actually say about you. Drop rank positions from AI reports entirely. For more on what a proper programme should report, see our guide to what AI SEO services should deliver.

What you can stop paying for

Google’s July 2026 guidance includes a list of GEO “hacks” you can ignore for Google Search. In short:

• llms.txt files and special AI markup. Google Search doesn’t use them. They won’t hurt, and other services may read them, but they won’t help your Google visibility.
• “Chunking” content into tiny pieces. Google says its systems understand multiple topics on one page.
• Rewriting content just for AI. AI systems understand synonyms, so you don’t need every long-tail variation.
• Inauthentic mentions. See point 9.
• Overfocusing on structured data. Useful for rich results, but no special schema is required for AI features.

That’s Google’s view of Google, not of ChatGPT or Perplexity, and it’s fair to treat any platform’s description of itself with some scepticism. But if an agency’s GEO proposal is mostly the items above, ask what they’re doing about points 1 to 11.

FAQ

What is GEO in marketing?

Generative engine optimisation (GEO) is the practice of getting a brand retrieved, cited and accurately described in AI-generated answers from tools like ChatGPT, Google AI Overviews, Gemini, Perplexity and Copilot. It combines technical access, content that’s easy to cite, and off-site reputation work, and it builds on traditional SEO rather than replacing it.

Is GEO different from SEO?

Partly. Google says optimising for its AI features is still SEO, and most AI answers are built from pages that already rank. The differences are in the extras: AI-specific crawlers, off-site mentions, brand accuracy, and measurement by visibility share rather than rank. In practice, the best results come from running both together.

Does llms.txt help with AI visibility?

Not for Google. Google states that Google Search, including its AI features, doesn’t use llms.txt or similar files. Some other AI tools may read them, so adding one is harmless, but it shouldn’t be the centrepiece of a GEO strategy or a reason to pay an agency.

Should I block AI crawlers?

Block training crawlers if your business has a reason to, but think carefully before blocking search crawlers. For example, blocking OpenAI’s OAI-SearchBot removes you from ChatGPT search answers, while blocking GPTBot only signals that your content shouldn’t be used for training. Check your CDN settings too.

How do you measure GEO results?

Use a fixed set of prompts, run each many times, and track how often your brand appears, how it’s described and who appears alongside you. Add first-party data from Google Search Console’s Generative AI performance report and Bing Webmaster Tools’ AI Performance dashboard. Ignore anyone reporting a single “ranking position” in ChatGPT.

Where to start

If you only do three things this month, make them the unglamorous ones. Check that AI search crawlers can reach you through your CDN and firewall. Fetch your key pages with JavaScript off and see what’s actually there. Then ask the major assistants what they think your business is, and write down what’s wrong.

Most of the GEO tactics agencies miss aren’t clever. They’re just the bits of the job that happen outside a content brief. If you’d like one team to cover the technical fixes and the digital PR, take a look at our GEO services or book a call.