Nine things in a GEO pitch that should end the meeting

We sell GEO, so read this as a seller telling you how to catch a seller. Nine things end the meeting for us: a guarantee of placement, a 40 percent lift figure, llms.txt on the invoice, schema sold as the reason you get cited, one blended visibility score, Reddit seeding described as outreach, no stated measurement method, Google and ChatGPT treated as one surface, and a refusal to say what they will not do. Each is either a claim the evidence contradicts, a deliverable that does nothing, or a way of hiding whether the work worked. Any one is enough. Two or three and you should be asking what else in the deck was made up.

The nine things in a GEO pitch that should end the meeting

1. Guaranteed placement

Nobody can guarantee that ChatGPT, Perplexity or Google's AI Overviews will name your business. There is no submission form and no index status page. Google's own guidance says there are no requirements beyond being indexed and eligible for a snippet. The models are non-deterministic, so the same prompt gives different answers on different runs.

Then there is the contract problem. In July 2026 OpenAI signed a non-exclusive licensing deal with Yelp. An independent test by Ben Fisher the following month, across 2,880 prompts, found Yelp grounding in about 96 percent of ChatGPT local runs. Local recommendations moved because two companies signed a document. An agency that guarantees placement is guaranteeing the behaviour of systems it does not control and cannot see into.

2. The 40 percent number

You will hear a visibility lift figure, and 40 percent is the one that travels. Ask for the study. Ask what was measured, across how many prompts, how many times each prompt was run, and on which surfaces.

Prompt sampling needs 12 or more repetitions per prompt before the result means anything. Five repetitions gives a margin of error around plus or minus 27 points. A lift figure with no sample size behind it sits inside that margin, and the person quoting it usually does not know.

3. llms.txt as a deliverable

Google states that Google Search ignores llms.txt. No major provider has committed to reading it for production answers. It is a documentation convention aimed at coding agents, and we have written up what it is for. If it appears on a proposal as a line item, you are being invoiced for a text file, and the agency either knows that or has not checked.

4. Schema sold as a citation tactic

Ahrefs ran a matched difference-in-differences study: 1,885 pages that added JSON-LD against 4,000 controls. AI Overview citations fell 4.6 percent, which was statistically significant. AI Mode and ChatGPT moved 2.4 and 2.2 percent, indistinguishable from zero. FAQ rich results ended in Google Search on 7 May 2026. The speakable property never left a news-only beta.

Schema is still worth having for conventional rich results and for telling machines which entity you are. We say as much elsewhere. Sold as the thing that gets you cited, it is a tell.

5. One blended visibility score

A single number that rolls ChatGPT, Perplexity, Claude and Google into "your AI visibility" hides everything you would want to know. The surfaces behave differently, they change on different schedules, and a move on one gets averaged away by the others.

It also hides the gap between being cited and being named. Semrush's ghost citation study found 61.7 percent of citations produced no brand mention. ChatGPT cited a source 87 percent of the time and named a brand 20.7 percent of the time. A score that counts a link to your page as visibility counts you as seen when nobody read your name.

6. Reddit seeding described as outreach

The evidence on mentions is real. Across 75,000 brands, Ahrefs found branded web mentions correlated with AI Overview visibility at 0.664, against 0.218 for backlinks. Muck Rack, across more than 25 million citations, put earned media at 84 percent of AI citations and paid or advertorial placement at 0.3 percent. Both are correlational, and both point the same way.

Agencies read that and propose to manufacture mentions: accounts dropping your name into Reddit threads, forum answers and comment sections. That is not earned media. It is astroturfing with your brand name attached when it gets found, and the platforms treat it as spam. Real outreach is a guide, a list or a journalist choosing to include you. It is slower, and it is what the correlations point at.

7. No stated method

"We will track your AI visibility" is not a method. A method names the prompt set, how many prompts, how many runs per prompt, which surfaces, what location and account state, and how often. It says what counts as a mention versus a citation. We have set out what a real tracking method looks like. If the agency cannot describe theirs in a paragraph, they are going to run a prompt once a month and send you a screenshot.

8. No separation of Google and ChatGPT

"AI search" is not one thing, and a pitch that treats it as one has not read a crawler log.

Vercel and MERJ analysed more than 500 million fetches and found no major AI crawler executed JavaScript. GPTBot, ClaudeBot and PerplexityBot fetch the files and never run them. Only Google's surfaces and Copilot render. So a site that builds in the browser can be fine in AI Overviews and invisible in ChatGPT, and a blended report will never tell you. The robots.txt guide covers which token governs which surface.

9. Refusal to say what they will not do

Ask it directly: what will you not do? A serious answer is specific. Ours is that we will not guarantee placement, will not put llms.txt on an invoice, will not seed mentions on platforms you do not own, and will not report one blended score. An agency that answers "we do whatever it takes" does nothing in particular, and you will find that out once the retainer starts.

What a pitch you can sit through looks like

It names the mechanical work first: server-rendered HTML, answering crawlers unblocked, 404s and redirect chains fixed. AI crawlers 404 at around 35 percent for ChatGPT and Claude against about 8 percent for Googlebot, per Vercel and MERJ, so this is not filler. It rewrites the paragraph after your H1, because that is where ChatGPT takes its stored snippet from.

It brings a fixed prompt set, run enough times, reported per surface. It points you to Bing Webmaster Tools' AI Performance report for Copilot citations and grounding queries, and to Search Console's Generative AI report for impressions.

And it is honest about scale. Conductor's benchmark puts AI referrals at around 1 percent of traffic. Ahrefs reported 0.5 percent of its own traffic from AI and 12.1 percent of signups, which is the more interesting number and also the harder one to attribute, because many AI referrals arrive with no referrer and land as Direct. A pitch that promises to show you the traffic is promising a report the data cannot support.

Print the nine and take them into the call. Tick each one as it comes up. Three ticks and the meeting is over, and you have kept the retainer.