The Multi-Engine GEO Pitch Is Right. Almost Every Number In It Is Wrong.

A newsletter landed in my inbox this morning from an SEO agency I’ve never worked with. The subject line had my first name in it and a promise attached: why you should be optimizing for Claude AND ChatGPT in 2026 (today).

I read it over coffee. The thesis is correct. I have been arguing a version of it on this site for eighteen months.

Then I did what I do with every vendor claim that crosses my desk — I went looking for the primary sources.

Ninety minutes later I had five headline statistics traced back to origin. One was attributed to a research firm that did not publish it. One described a market mechanism that has since inverted. One was a single agency panel presented as a market universe. One had no locatable source at all. One was a real number from a real study — that the study’s own authors have superseded twice since.

None of this makes the argument wrong. That is the uncomfortable part, and it is the reason this post exists.

I have spent thirty years watching enterprise software vendors sell directionally correct theses on top of mechanically expired evidence. It ends the same way every time.

Why This Pattern Bothers Me More Than It Probably Should

In 1993 I was at KnowledgeWare, the Atlanta CASE tools company founded by James Martin) and run by Fran Tarkenton, selling model-driven development into Fortune 500 IT organizations. The pitch was airtight: application backlogs were exploding, hand-coding didn’t scale, model-driven development was the future. Every word of that was true. KnowledgeWare was sold to Sterling Software in 1994 after the bottom fell out of the category in about thirty-six months.

What killed CASE wasn’t the thesis. It was that the vendor evidence — the productivity multiples, the defect-reduction curves, the reference accounts — had been assembled under conditions that no longer existed by the time buyers committed capital against it. When the mechanism underneath a correct thesis stops working, the thesis becomes an expensive way to be right.

That is precisely what is happening in generative engine optimization right now. The GEO thesis is sound. The GEO evidence base is being recycled, relabeled, and resold at a rate that should alarm anyone spending seed capital against it.

Claim 1: The 680 Million Citation Study

The newsletter cites a 680 million citation index published by 5W Communications in May 2026 as the basis for its per-engine source breakdown.

The dataset is real. The attribution is not.

The 680-million-citation analysis belongs to Profound, and it covers August 2024 through June 2025. It has been in open circulation for roughly a year — Leads Now AI and QuickSEO both cite it correctly, with dates. It was not published by 5W Communications, and it was not published in May 2026. A study whose collection window closed thirteen months ago is being presented as fresh 2026 market intelligence.

There is a second, subtler problem. The newsletter reports that ChatGPT pulls 47.9% of citations from Wikipedia. Profound’s actual finding is that Wikipedia represents 47.9% of citations within ChatGPT’s top ten most-cited sources — a share-of-leaders metric. Wikipedia’s share of ChatGPT’s total citation volume is approximately 7.8%.

Those are not the same number and they do not support the same strategy. Read 47.9% as “half of ChatGPT’s citations come from Wikipedia” and you conclude the citation market is a near-monopoly you can’t enter. The real distribution says close to the opposite: Profound’s separate 730,000-conversation study, summarized by Pressonify, found the top ten domains capture only about 12% of citations. The long tail is where a seed-stage company can actually compete.

A metric quoted without its denominator is not evidence. It is decoration.

Claim 2: The Market Share Table

The newsletter’s centerpiece is a clean, memorable set of figures: ChatGPT fell from 89% of B2B AI referrals to 62.6%, with Claude at 18.5%, Gemini at 10.6%, and Perplexity at 7.3%.

Those numbers are traceable. They come from Goodie’s 2026 AI Search Traffic Report, published May 2026 — a Wave 2 study built on GA4 referrer data from an anonymized brand panel, triangulated against SimilarWeb, and brand-averaged so each site counts equally. Wave 1 measured roughly 2.8 million AI referral sessions across 41 brand sites.

It is legitimate research, honestly documented by its authors. It is also one panel, and its methodology is doing enormous work. Here is the same question from three other vantage points:

Source	What it measures	ChatGPT	Claude
Goodie (May 2026)	B2B AI referrals, ~41-brand panel, brand-averaged	62.6%	18.5%
Statcounter (Mar–Apr 2026)	Global AI chatbot referrals, all verticals	76.9–78.2%	2.66–2.91%
Previsible (Jul 2026)	Trackable LLM referral sessions	92.4%	overtook Perplexity Mar 2026

Claude is either 18.5% of the market or 2.9% of it depending on which denominator you accept. Both readings are defensible. Goodie’s figure is B2B-specific and brand-averaged, which structurally inflates engines that over-index on technical and professional audiences — exactly what Claude does. Statcounter’s is global and all-vertical, which structurally deflates it. Statcounter CEO Aodhan Cullen explicitly cautioned that Claude’s March spike partially retraced within the same month.

The newsletter presents one panel’s B2B-weighted figure as the state of the market, with no methodology note and no attribution.

And then there is the number that quietly undermines all three columns. The Digital Bloom’s February 2026 analysis of 446,405 visits found that 70.6% of AI traffic arrived with no referrer header at all — invisible to standard GA4 attribution and misfiled as direct traffic.

Every referral-share statistic in this category, including the ones I just tabulated, measures the visible minority. We are arguing about the composition of roughly three sessions in ten and calling it market share.

If you cannot see 70% of a channel, you do not have a market share number. You have a sampling artifact with a decimal point.

Claim 3: The AI Overviews Mechanic Has Inverted

This is the claim that matters most, because the newsletter converts it directly into a tactic.

The argument runs: 76.1% of AI Overview citations come from URLs already ranking in Google’s top ten, therefore the retrieval mechanic is direct — rank on Google, get cited in AI Overviews.

That figure is real and I can name its origin precisely. It comes from Ahrefs, authored by Louise Linehan with Xibeijia Guan, analyzing 1.9 million citations across 1 million AI Overviews. It found 76.10% of cited pages ranking in the top ten, 9.50% in positions 11–100, and 14.40% ranking nowhere.

It was accurate. It is now roughly two years old, and the relationship it described has collapsed.

BrightEdge reported 54% overlap in October 2025. By February 2026, BrightEdge’s one-year-mark study measured approximately 17% — and noted the figure had been flat for months, with roughly five of every six AI Overview citations pulling from content not on page one. Ahrefs’ own follow-up, covered by Search Engine Journal and DesignRush, put it near 38%. ALM Corp’s cross-methodology analysis documents the full arc from 76% to the 17–38% range.

The driver is query fan-out — Google’s AI decomposes a question into sub-queries and evaluates content against prompts the site owner never tracked. Google promoted Gemini 3 to global default for AI Overviews on January 27, 2026; SE Ranking’s post-upgrade analysis, summarized by SEOScaleUp, found Gemini 3 replaced approximately 42% of previously cited domains and returned about 32% more source URLs per response.

So the newsletter’s recommended tactic — audit your top-ten keywords, find the ones triggering AI Overviews, restructure those pages — rests on a mechanism that has degraded by half to three-quarters since the number was true. Top-ten ranking is now a weak predictor of AI Overview citation.

Two further corrections in the same section.

Prevalence. The newsletter says AI Overviews appear on 25% of searches, up from 6.49% in January 2025. The baseline is right. The current figure traces to Semrush’s July 2025 peak of about 24.6%, which pulled back to roughly 15.7% by November 2025, while BrightEdge’s February 2026 tracking put presence near 48–50%. Slate’s roundup frames it correctly: 15–25% in conservative mixed-intent datasets, 48–50% in informational-heavy panels. It depends entirely on the query set, and quoting one number without the query set is meaningless.

CTR. The newsletter cites a drop from 1.41% to 0.64%. That is a genuine Seer Interactive figure — from their January 2025 original study. Seer has updated it twice since. Their September 2025 update revised the range to 1.76% → 0.61% across 3,119 queries and 42 organizations, covered by Search Engine Land. Then their 2026 update, authored by Tracy McDonald across 2.43 billion impressions and 5.47 million queries, found organic CTR on AIO-present queries rebounding 85% in the first two months of 2026, back to roughly 2.4%. Seer says so plainly: their own 2025 model predicted continued decline, and their 2026 data disrupted that assumption.

The newsletter is quoting an eighteen-month-old floor from a study whose authors have publicly retired the trendline.

And here is the finding it omitted — the only genuinely actionable number in the entire AI Overview literature. Per Seer, brands cited inside an AI Overview earn approximately 35% more organic clicks and 91% more paid clicks than uncited brands on the same query. Later per-impression analysis puts the cited-versus-uncited advantage nearer 120%.

Citation is the variable. Ranking is no longer a reliable path to it.

Claim 4: The Claude Structural Claim

The newsletter asserts Claude is 30% more likely to cite bullet-pointed pages than other engines, and builds a rewrite recommendation on it.

I could not locate a source. Not in Profound’s index, not in ZipTie’s per-platform retrieval analysis, not in Discovered Labs’ citation-pattern work, not in any primary study I could reach.

What the same Profound dataset actually reports, per ZipTie, is that Claude’s top-cited category is blogs, at 43.8% — a content-type preference, not a formatting preference. That distinction has budget consequences. “Add bullets to your comparison pages” is a two-hour formatting task. “Publish sustained original analytical writing under a named author” is a program. The newsletter recommends the first; the data supports the second.

Worth setting alongside it: SearchSignal’s benchmark finds Anthropic’s Claude crawls roughly 38,066 pages for every referral it sends, against ChatGPT’s 1,091. Claude reads vastly more than it forwards. Optimizing for Claude is substantially a corpus problem rather than a click-capture problem — which is the premise behind our 7-tier LLM Training Data Audit and the 42-source training data matrix we published in May.

The directional claim about Claude traffic quality does hold, and it deserves credit. The Digital Bloom measured a 16.8% conversion rate on Claude-referred sessions — highest of any platform — against ChatGPT at 14.2–15.9%, Perplexity at 10.5%, Gemini at 3.0%, and a Google organic baseline of 1.76–2.8%. Claude traffic is genuinely the most valuable per session in B2B. The newsletter’s “5–10x versus ChatGPT” framing is not what that data says. We made the same argument about unsourced conversion multiples in The 30-40% LLM Conversion Rate Claim Is One Company’s Anecdote.

Claim 5: The Bing Dependency

The newsletter states 92% of ChatGPT retrieval runs through Bing, and recommends Bing rankings as the lever.

Bing matters. It is not 92%, and the trend runs against the claim. Profound’s tracking, documented by Pressonify, shows ChatGPT’s Bing alignment falling from 26% to 8% while Google alignment rose from 12% to 33%. OpenAI now operates its own crawler, OAI-SearchBot, and ThePlanetTools’ analysis documents that Bing dependence has been declining. Ahrefs found only about 10% of ChatGPT-cited URLs ranking in Google’s top ten.

The operational takeaway survives — allow OAI-SearchBot in robots.txt, verify Bing indexation, don’t let a crawler block cost you a retrieval path. Digital Applied’s case study on a site losing 90,000+ ChatGPT citations after a Bing deindexing shows the tail risk is real. The stated mechanism is not.

The Three Questions I Run on Every GEO Statistic

1.  Who published it, and when was the data *collected*? Not published — collected. Profound’s index closed in June 2025. Ahrefs’ 76% study predates query fan-out. Anything built on either describes a pre-Gemini-3 retrieval environment. Publication dates in this category are marketing artifacts.

2.  What is the denominator? Share-of-top-ten is not share-of-total. B2B brand-averaged is not global. Trackable referrals are not referrals. A number without its denominator cannot be compared to the number beside it in the same table.

3.  Has the mechanism been retested since the models changed? Gemini 3 became the AI Overviews default on January 27, 2026 and replaced ~42% of cited domains. ChatGPT’s index composition shifted materially across 2025–2026. Any causal claim about retrieval predating those changes needs re-verification before it becomes a budget line.

If a vendor cannot answer all three about their own headline statistic, you are not buying research. You are buying confidence.

What Founders Should Actually Do

Pre-seed and seed. Do not buy a multi-engine program off a market-share slide. Establish your own baseline. Run the twenty prompts your buyers would run across ChatGPT, Claude, Gemini, Grok, and Perplexity, log whether you appear, repeat monthly. Your panel is you. Seer’s own conclusion after 2.43 billion impressions is that aggregate benchmarks mislead and your own data is the strategic input. The mechanics are in The AI Dark Funnel.

Fix attribution before content. GA4’s native AI Assistant channel does not include Perplexity, and Claude’s inclusion is unconfirmed. Build a custom regex channel group or you will spend a quarter optimizing against a dashboard that cannot see the traffic. Given the 70.6% referrer gap, add a “how did you hear about us” field to your demo form — it will outperform your analytics stack.

Invest in what is engine-agnostic. Across every credible study, three properties predict citation everywhere: ungated, structurally clean, original. Gated whitepapers generate zero signal across all five models — documented in our LLM training data due diligence guide and in Your Gartner Placement Is Invisible to Every Major AI. That is a publishing strategy, not an engine strategy, and it is the only part that doesn’t expire when a model version ships.

Approaching a transaction. Citation footprint is now a diligence line. Buyers are beginning to price AI search visibility as a transferable customer-acquisition asset — the argument I made in The Answer Economy and The AI Search Visibility Audit Your Deal Room Is Missing.

The Part That Should Bother You

The newsletter’s core recommendation — build visibility across every engine, because the retrieval pipelines genuinely differ and optimizing for one does not transfer — is correct. I would sign my name to it.

But an argument this good does not need borrowed authority, a misattributed study, and an eighteen-month-old CTR floor dressed in this quarter’s date. That it was built that way anyway tells you how the GEO services market is currently competing: on the confidence of the claim rather than the durability of the mechanism.

I saw this movie at KnowledgeWare. The vendors with the most compelling evidence decks were not the ones with the best products. They were the ones who had stopped checking whether their evidence was still true.

Your GEO strategy should be multi-engine. Start by auditing the evidence you are building it on.

DevelopmentCorporate LLC runs structured LLM Training Data Audits and 90-day GEO implementation plans for pre-seed through Series B B2B SaaS companies. Every claim in our audits is traced to a primary source with a collection date — see our published research reports. To see where your company actually stands across ChatGPT, Claude, Gemini, Grok, and Perplexity, book a discovery call and we will run three live test queries for your category on the call.