We Asked Google's AI Which GEO Tool to Use. The Funded Ones Were Absent.

We asked Google’s AI Overview, six times, in the United States and in English: what is the best tool to track my brand’s visibility in ChatGPT?

Sixteen distinct domains were cited. Five appeared in every run. Not one of them belonged to a company that has raised serious money in this category.

No Profound, which has raised more than $155 million at a $1 billion valuation. No AirOps, with roughly $118 million raised. No Scrunch, acquired by Sitecore. No Peec AI, no Evertune, no Semrush. The answer named tools most people in this industry have never heard of.

Disclosure: EchoWi is our product and it competes in this category, so this is a compilation by a company with an interest in the result. Everything needed to repeat it against us is here: the question, the surface, the market, the language, the run count and the dates, all stated below, and the rows are in our measurement register.

The short version

1. Sixteen domains cited across six runs. Five appeared every time. Eight appeared exactly once.
2. The best-funded vendors in the category did not appear at all, across six runs, on the query that describes their product.
3. Two of the five stable domains are not tools. Reddit and YouTube.
4. Two are tools almost nobody in the industry discusses: siftly.ai and alhena.ai.
5. Funding does not buy AI visibility. That is the finding, and it is uncomfortable for everyone selling in this space, us included.

Method

Three method notes that decide whether this means anything.

The market is explicit. Measurement tools default to the United States and English. Left unset, a question in another language gets answered by the American market and the report does not say so. We set it deliberately here, and we set Spain and Spanish deliberately when we put the same question to three markets.

The cache is bypassed. If runs reuse a stored answer they return an identical list and report perfect consistency about something volatile. That measures the opposite of what you want.

Runs are serial. Firing them together hits the same session within milliseconds, correlates the samples, and hides the variance you are looking for. It is slower for a reason.

And the honest limitation: we asked about ChatGPT, but we measured Google’s AI Overview. This is what Google’s AI answers when someone researches the question, which is a real and common path. It is not what ChatGPT itself would answer.

The result

Eight domains appeared exactly once. A single check would have presented one of those with the same confidence as siftly.ai, which appeared every time.

Who was missing, and what it cost them

This is the part worth sitting with. The companies absent from all six runs have raised, between them, well over $200 million:

Meanwhile siftly.ai and alhena.ai, which have no comparable public funding and appear in no major comparison article we know of, were cited in every single run.

We are not claiming those two are better products. This measurement says nothing about quality. What it says is that the mechanism deciding who gets recommended is not the mechanism deciding who gets funded, and a lot of category marketing is built on the assumption that they are the same.

Why this might be happening

Four explanations, ordered by how well the available evidence supports them, and none of them proven by this measurement alone.

1. Retrieval favours pages that answer the exact question. The strongest finding across the research is that relevance to the derived sub-questions dominates. A vendor homepage sells a product. A page titled around the literal question answers it. Ahrefs found that similarity between a page title and the model’s derived sub-queries separated cited from uncited URLs across 1.4 million prompts.

2. Third-party and community sources carry weight. Reddit and YouTube in every run is not an accident. Neither sells a tool, and both contain people answering the question directly.

3. Funding buys sales capacity, not retrieval. A Series B pays for account executives and enterprise features. Neither is visible to a retrieval layer.

4. The category is young and the index is thin. With few genuinely useful pages answering this question, small sites that answer it plainly can outrank large ones that market around it.

What this does not show: that any of these is causal. A July 2026 review of 45 GEO studies found no technique with a demonstrated causal, stable, cross-platform effect. This is one question, one surface, one market, six runs.

What it means if you are choosing a tool

• Do not treat an AI recommendation as a shortlist. Five of the sixteen domains here were stable and two of those are not products. An assistant’s answer to a buying question reflects who wrote a good page, not who built a good product.
• Do not treat absence as failure either. Six serious companies scored zero. If your own brand is missing from an AI answer, that is a content and retrieval problem, not proof that your product is worse.
• Check the market you actually sell in. In Spain and in France the same question returned an almost completely different set, including a Spanish tool that appears in no English comparison.

What it means if you are trying to be recommended

The uncomfortable implication is also the actionable one. The tools winning here are winning with content, not with capital. They published something that answers the question directly, and the retrieval layer does not know or care what their valuation is.

That is consistent with what has controlled support: in the Princeton GEO-bench experiment, adding quotations, statistics and cited sources raised a source’s share of the answer by 41%, 30% and 27%. None of those edits require funding.

Nobody funded appeared, and that is the point

EchoWi sells exactly what this question asks for, so this is a measurement run by an interested party. The way to handle that is not to claim neutrality, it is to state the conflict and make the run repeatable: the question, the surface, the market, the language, the run count and the date are all above, nothing in the catalogue was exempted, and you can put the same question to the same surface yourself.

The result worth publishing is not about any one vendor, it is the shape of the answer. Profound, AirOps, Scrunch, Peec AI, Evertune and Semrush are the funded names in this category, and not one of them appeared in any of the six runs. Between them they have raised or are worth hundreds of millions of dollars to be found in AI answers, and the two domains that were cited in every single run, Reddit and YouTube, sell nothing in this category at all.

That is uncomfortable for everyone selling here, which is why it is worth the space. A category can be well funded, well marketed and entirely absent from the answer it is built to win, and the money is visibly not what closes the gap.

Measured again, and it moved

We re-ran the same question on 7 August 2026, four runs, same surface, same market, counting citations only.

More domains, less agreement. Only three held across all four runs, and fourteen of the twenty-three appeared once and never again.

That is worth stating plainly because we have re-run two other questions this week and both went the other way. Asking which website is best for booking flights returned six cited domains, all six in every run. Asking which noise cancelling headphones to buy returned four, all four in every run. Those categories are settled. This one is not.

The difference is not the method, which was identical in all three. It is the category. Flights and headphones have been reviewed by established publications for twenty years, so retrieval has obvious places to go. “Best tool to track brand visibility in ChatGPT” has almost no settled literature, so each run assembles a different set from a long tail of vendor blogs, and fourteen of them appeared exactly once.

Which means the answer to “how many runs do I need” is not a constant. In a settled category, three runs tell you almost everything. In this one, three runs would have shown you three domains that repeat and hidden the fourteen that do not.

The funded vendors were absent from one surface, not from the market

Everything above is Google AI Overview. On 7 August 2026 we asked the same class of buying question on all four surfaces at once, United States and English set explicitly: which AI visibility tool should I use to track how my brand appears in ChatGPT?

AI Overview returned an upstream server error rather than an answer, so three surfaces replied and one failed. That is our instrument failing, not Google staying silent, and we are not going to report it as a finding either way.

The three that answered contradict the headline of this article, and they should be allowed to.

Profound was named by all three. ChatGPT put it first, Gemini listed it under enterprise, AI Mode opened with it. The vendor that was absent from six AI Overview runs is the vendor the other three surfaces reach for. So the absence we measured was a property of one surface on one day, not a property of the market, and the sentence at the top of this article should be read with that attached.

Two more of the funded names appear too: Peec AI on AI Mode, Otterly on Gemini and AI Mode.

The citations still do not match the recommendations

The pattern this article found holds, and holds harder. Across the three surfaces the answers rest on 21 cited domains. One appears on two surfaces. None appears on all three.

More to the point, the domains cited are mostly not the vendors recommended. AI Mode names five tools and cites a dozen sources, and the sources are third-party listicles: an agency’s blog, a hosting company’s blog, a martech vendor’s blog, a consultant’s site, a business directory. One of them is published by a tool that sits in our own catalogue, ranking its competitors. Gemini’s citations concentrate on two vendor blogs, one of which it cites six times.

So the recommendation and the evidence come from different places. The vendor gets named; a third party gets the citation. If you are a vendor trying to be recommended, that gap is the whole job, and it is not on your own site.

What the demand says about the same names

We checked the search side the same day, market and language explicit, with close-variant disaggregation on. “Profound ai” is 6,600 searches a month in the United States, and we read its results page to be sure the volume was the vendor’s rather than a homonym: seventeen of the eighteen organic results are the AI visibility company. “Peec ai” is 2,400.

Those are the two largest verified demand figures for any tool in our catalogue, and they now belong to the two vendors most often named across surfaces. On AI Overview alone, neither appeared at all.

Three named tools we do not carry

The three surfaces named three products that are not in our verified catalogue: Hall, GRRO and PromptRush. We measured each one before deciding, and none returns measurable search demand in the United States, so they are recorded here rather than catalogued. A tool an assistant names is not automatically a tool anyone is looking for, and the two facts are worth keeping apart.

This is one observation. One prompt, one day, one market, one pass on each surface that answered.

Common Questions About This Measurement

Which tools did Google’s AI Overview recommend?

Across six runs, the domains cited every time were reddit.com, seranking.com, youtube.com, siftly.ai and alhena.ai. Of those, only the last three are products, and only two are AI visibility tools specifically.

Why did the well-funded vendors not appear?

This measurement cannot say for certain. The most likely explanation, and the one best supported by published research, is that retrieval favours pages that directly answer the question asked, and vendor marketing pages are written to sell rather than to answer. Funding does not affect retrieval.

Is six runs enough?

It is enough to separate stable citations from coincidences, which is the purpose here: five domains appeared every time and eight appeared once. It is not enough to publish a precise rate with error bars. Every run is billed and executed serially, so sample size has a real cost.

Does this mean siftly.ai and alhena.ai are the best tools?

No, and we would push back on anyone who read it that way. This measures citation, not quality. No controlled accuracy comparison of tools in this category has been published by anyone, including us.

Why measure Google AI Overview instead of ChatGPT?

Because it is one common way people research this question, and because it is directly measurable. It is a genuine limitation: what Google’s AI says about ChatGPT tools is not what ChatGPT says. A fuller study would cover several surfaces.

Can I reproduce this?

Yes, and that is the intent. The question, surface, market, language, run count and date are all above. If your result differs, it is because the model changed or because the market was not set explicitly.

Did any funded vendor appear?

No. Profound, AirOps, Scrunch, Peec AI, Evertune and Semrush were absent from all six runs, and the two domains cited every time, Reddit and YouTube, sell nothing in this category. We sell here too, so this is a count published by a company with an interest in it, and the six runs, the surface and the date are stated so anyone can redo it against us.

Where this leaves you

The single most useful thing in this data is not the list of winners. It is that six companies holding more than $200 million between them appeared zero times on the query that describes their own product.

If capital determined AI visibility, that could not happen. It happened across six runs.

Next: what GEO is and what the research supports, and the tool comparison with verified prices.