GEO engine: 23 percent of news citations came from Grok, and our registry reads none of them

What is a GEO engine, and how many of them are there?

There is no register of generative engines and no definition that decides membership, so any count is a count of what somebody chose to monitor. Goodie's study states that it monitors brand visibility across ChatGPT, Gemini, Claude, Perplexity, AI Overviews, AI Mode, Copilot, Grok, DeepSeek, Meta AI and Amazon Rufus. That is eleven, and the interesting part is not the number but how unlike each other the members are.

Four of them are chat products with a retrieval tool attached. Two of them, AI Overviews and AI Mode, are features of a search results page rather than products you open. One of them, Amazon Rufus, is an assistant inside a shopping app, where the corpus it answers from is a retail catalogue rather than the open web. Grouping those under one word is convenient, and it is also why two vendors can report different numbers for the same brand without either being wrong: they counted different sets of machines.

The distinction that matters for measurement is not the product type but what each surface hands back. An engine that returns the URLs it consulted can be measured for citations. An engine that answers from what a model retained during training returns nothing to count, and a tool asking it is measuring memory rather than the live web. Both produce a fluent paragraph naming brands, and nothing in the shape of the answer separates them, which is the argument our AI visibility entry makes and the reason a report should say which kind it asked.

Two of the eleven are separately worth pulling apart, because they are often treated as one thing. Google's AI Overviews and Google's Gemini app are different surfaces with different retrieval paths and different citation behaviour, which is why we argued that a Gemini citation tool is not an AI Overview tool rather than a cheaper version of one. The same care applies to the older label: answer engine optimization grew up around question answering surfaces, and a good deal of what it describes still holds, but the set of surfaces it was written for is not the set above. Before any of this matters, a page has to be fetchable and readable at all, which is the other layer entirely and the one most sites fail first.

Where the news citations landed, according to one 31 million citation panel

The study's own summary of the collection is that between October 2025 and July 2026 Goodie recorded 31 million AI citations, and that news makes up between 1.3 and 1.9 percent of all citations in any given month. Both halves of that sentence are load bearing. The engine shares below are shares of the news slice, and the news slice is a small minority of everything these surfaces cite, so nothing here describes the whole citation surface for an ordinary business.

Within that slice, the study reports Grok as the single biggest citer of news at 23 percent of all earned citations, then AI Overviews at 19.5 percent, Gemini at 13.5, ChatGPT at 11.9, AI Mode at 10.6, DeepSeek at 7.9 and Claude at 5.1. Those are seven of the eleven surfaces it monitors, and the seven published shares sum to 91.5, which leaves 8.5 points spread across the four surfaces the study does not give a news share for in what we could read: Perplexity, Copilot, Meta AI and Amazon Rufus.

The shape is the finding. No engine holds a quarter of the news citations, the top two are eleven points apart, and the fourth largest is the one most people would have named first. A visibility number quoted from one engine is therefore a number about somewhere between 5 and 23 percent of the citations in this slice, which is the same problem in a different form as the one where more AI citations did not mean more of your page in the answer.

The publisher side of the same study is the sharper number. It reports that Forbes holds 33 percent of all news citations, that the top five publishers take 66 percent and the top ten take 84, and that of 495,000 citations to 37 major news domains, 34 recorded any citations at all. Concentration like that is consistent with what has been measured elsewhere about who gets cited: our reading of publisher deals found 10.2 ChatGPT citations per page against 6.9 for licensed publishers, and separately that citations reached 6.8 percent of ChatGPT prompts and the visit landed on the homepage. For a business that is not a newsroom, the more useful published pattern is that most cited domains are other companies rather than the brand being asked about.

Three of the seven named engines are Google, and they add to 43.6 percent

Add the three Google surfaces in that table and the arithmetic is 19.5 for AI Overviews plus 10.6 for AI Mode plus 13.5 for Gemini, which is 43.6 percent of the news citations in the study's sample. That sum is ours, computed from their published figures, and it is worth stating because it is invisible in a per-engine ranking where Google appears three times under three names.

It would be wrong to read that as one system with three faces. Two of the three are documented as sharing an eligibility gate: Google's AI features documentation, carrying Last updated 2025-12-10 UTC, says that AI Overviews and AI Mode surface relevant links to help people find the information they are looking for quickly and reliably, and that to be eligible to be shown as a supporting link in AI Overviews or AI Mode, a page must be indexed and eligible to be shown in Google Search with a snippet. That is a shared gate for two of the surfaces. The Gemini app is not named in that requirement.

So a site owner working on Google has one control that plausibly moves two of the three surfaces, and we have measured how rarely that control is even touched: across 81 home pages that answered on 6 September 2026, none had blocked their own snippet. The harder half is that the fetching side of these features is not a crawler anyone can point at with confidence, which is why we do not model an AI Overviews crawler and say so on the report rather than inventing one.

The same vendor arithmetic runs the other way at Microsoft. Copilot appears once in the study's surface list with no published share, and its grounding behaviour is unusually legible compared to Google's: Bing exposes the query it ran, which is why we could write that Bing names the grounding query and Google names none. Legibility and citation volume are different properties, and a vendor can be strong on one and absent on the other.

What this scanner asks, and the 33.6 points of news citations it cannot read

Set that study against our own registry and the gap is specific rather than rhetorical. Read out of core/src/engines.ts on 7 September 2026, Lantad ships 8 answer engine rows: an open-weight chain running on Cloudflare Workers AI, then Gemini, DeepSeek, Grok, Perplexity, ChatGPT, Claude, and Google AI Overviews captured from the results page. The registry names each of those by vendor rather than by the exact model called, which is an admin detail. Seven of those name a surface the study also monitors. The eighth, the open-weight chain, is not a consumer surface at all.

Of the seven, five are configured to retrieve live and return the sources they grounded in. Two are not. The DeepSeek and Grok rows both record retrieval as none and citations as false, and in the Grok case the registry comment states the reason plainly: xAI's Live Search is deliberately not wired because it bills per source used, which no flat per-call booking can honestly cover. That is a cost decision this product made, not a finding about xAI, and the distinction between a setting and a measurement is one we try hard not to blur. The full retrieval split is in 3 of the 8 engines we ask cannot search the web.

Put the two together and the inconvenient number falls out. The surface the study ranks first for news citations, Grok at 23 percent, is one this scanner asks without retrieval and reads no citations from. AI Mode at 10.6 percent is not asked at all. That is 33.6 points of the study's news citations that our citation numbers cannot see, against 50.0 points from the four surfaces where we both ask and read sources and where the study publishes a share: AI Overviews, Gemini, ChatGPT and Claude. Perplexity is a fifth such surface with no published share to add.

None of that is unique to us, and the category's shape is the reason: reading pages and asking models are two different machines sold under one name, which is what we found when ten AI visibility products were read at their own pages and three of them fetched anything. Coverage claims are worth checking against the vendor's own documentation on both halves. On the crawler half, the tokens are at least published: OpenAI's crawler documentation states that OAI-SearchBot is for search, that GPTBot is used to make their generative AI foundation models more useful and safe, that ChatGPT-User covers certain user actions in ChatGPT and Custom GPTs, and that OAI-AdsBot validates the safety of pages submitted as ads. Four tokens, four purposes, one vendor. Our own methodology page is where the equivalent disclosure for this product lives.

What to check before you trust any per engine visibility number

The practical value of a distribution like this is that it turns a vague worry into four questions you can put to any tool, ours included, and get a specific answer or a silence.

Which surfaces were asked, named individually. A tool that says it covers the major AI engines has told you nothing, because the set is unstable and the shares are not close to even. If the biggest single share in the news slice belongs to a surface a tool does not cover, that tool's total is not wrong so much as it is about a different population.

Whether each surface retrieved or remembered. A model answering with no search tool attached returns no sources, so any citation figure for it is either absent or derived from something else. This is where a report either states the retrieval mode per engine or quietly averages two different measurements together.

How many runs and how many prompts the number rests on. Answers move between identical runs, and we measured that AI visibility took seven runs per prompt to settle before the variance stopped mattering, and separately worked out how many prompts a measurement needs before it means anything. A single run against a single prompt is an anecdote with a percentage sign attached.

Whether the crawler half was checked at all. Being cited requires being readable, and blocking is per token rather than per vendor, which is why a robots.txt edit is a weaker instrument than it looks: a canary study found a block did not stop 12 of 18 AI chatbots from returning the content. Our AI crawler reference lists the tokens worth naming, and what AI says about a domain is the answer side of the same question. Neither is a substitute for the other, and a tool that sells one while implying both is the failure this category keeps repeating.