The Reddit summary of Demand-Genius’s GEO survey exposes a reporting problem, not proof that citations produce revenue. I chose it for this teardown because its headline figures, 77% reporting citations versus 32% reporting leads or pipeline, invite exactly that confusion.
The underlying report, Overwhelmed, Unsure, Unproven: Selling GEO in 2026, draws on 47 people at 46 B2B agencies surveyed in August and September 2026. But those two percentages come from just 22 agencies that disclosed what they report. They describe reporting practices, not revenue outcomes.
That distinction cuts both ways. Reporting citations does not establish that citations generate pipeline. Failing to report attributed pipeline does not establish that the work generates none. Before those numbers become another slide in an agency pitch, I want to take apart what they actually say.
First, inspect the sample and the seller
The report starts with 47 people at 46 B2B agencies. Three-quarters run or co-run their agency. Most agencies have fewer than 25 people, and most are UK-based. That is the base behind the findings about to get quoted at you.
My verdict on the sample: use it as directional evidence from mostly small, UK-based sellers of GEO, not an industry census. It tells me something useful about these respondents. It cannot stand in for every agency, every buyer or every commercial outcome.
Then read who pays for the lens. Demand-Genius discloses that it sells in this market. It describes rebuilding its own visibility tracking around buyer segments and clusters, and the report ends by inviting agencies to buy its research, technology and expertise.
That does not make the numbers false. It does mean the seller benefits when prompt tracking looks complicated enough to require its method. I read that commercial context alongside the findings, not as a reason to dismiss them and not as a footnote to forget.
Note the sample and the seller before quoting a single percentage.
Claim one: 77% report citations or mentions
Of the 22 agencies that disclosed their reporting, 77% report citations or mentions, 73% report share of voice and 68% report referral traffic. Only 36% report how the brand is described in answers.
That last number deserves more attention. A citation links to you as a source. A recommendation tells the buyer to consider you as a supplier. Those are different jobs. Counting them together hides the distinction a buyer actually cares about: did the answer use your information, or did it suggest buying from you?
I have written about being cited as a source without being recommended as a supplier. A report can look busy while leaving that question unanswered. Forty citations make a substantial PDF. They do not, by themselves, make a sales conversation.
My verdict: a rising citation count shows that the tracked answers cited you, not that a buyer chose you. Even the visibility claim needs its boundaries: which questions, which engines and which runs?
Before renewing, ask for the sentence around the citation. Was your brand:
- Recommended in answer to a buying question?
- Mentioned in a list of five options?
- Linked as background reading beneath an answer that recommended someone else?
If the agency cannot show the prompt, full answer and date, you bought a count without its context. Separate recommended, mentioned and merely linked before deciding what the count is worth.
Claim two: only 32% report leads or pipeline
The 32% leads-or-pipeline figure can also get read as an admission that GEO never pays. It is nothing of the sort. The survey asks what agencies report. It does not establish what their clients earn.
There is a real attribution problem here. Citation rate, share of voice and visibility scores can rise while revenue stays flat. Equally, a buyer can encounter a brand in an AI answer without leaving a neat trail from that answer to the CRM.
The plumbing helps explain the difficulty. Google counts AI Overviews and AI Mode clicks within Organic Search, while ChatGPT referrals can appear under their own source. And Pew found that AI-summary links were clicked in only 1% of visits where a summary appeared. A click may sit inside a broader channel, or there may be no click to follow.
None of that proves why a particular respondent does not report pipeline. It shows why “no attributed pipeline in the report” and “no commercial effect” are not interchangeable.
My verdict: missing pipeline reporting is a gap in the evidence, not a measured revenue result of zero. The reverse matters too: a measurement difficulty is not permission to claim revenue nobody can demonstrate.
If I am buying ongoing SEO and generative engine work on pipeline claims, I want three separate records:
- Observed traffic: identifiable AI-referred sessions, kept distinct from broader organic traffic wherever the reporting allows it.
- Buyer self-report: a “How did you hear about us?” field with AI options, carried into the CRM rather than buried in a form export.
- Commercial outcomes: leads and pipeline matched to those records, with the attribution limits stated.
A 30-to-90-day matching window provides a way to examine the connection. It does not turn a citation into proof of causation. Nor should self-reported discovery and tracked referrals disappear into one number that looks more certain than either record supports.
Ask what was observed, what was attributed and what remains unmeasured. Get all three in writing.
Claim three: 11 of 20 tool users do not fully understand prompt selection
This is the finding I would spend the most time on. Among agencies using a third-party visibility tool, 11 of 20 did not fully know how it decides which prompts to track.
Respondents who said they fully understood prompt selection rated their tool’s accuracy at 4.00 out of 5. Those who only roughly understood it rated accuracy at 1.67. That is a difference in reported confidence, not an independent accuracy test. Understanding the method and trusting the result travel together here; the survey does not establish which causes which.
The underlying problem still matters. The question set defines what the score covers before a single answer is read. A visibility score for questions buyers never ask can be perfectly consistent and commercially unhelpful. A score that changes because the question set changed tells a different story from one that changes on a fixed test.
My verdict: without the prompt list, I cannot judge what a visibility score represents. I do not need another decimal place. I need to see the questions.
Repeated runs matter too. SparkToro had 600 volunteers run 12 prompts 2,961 times across ChatGPT, Claude and Google AI. Its reported findings were that the same brand list appeared in two responses less than 1 in 100 times, and the same order about 1 in 1,000.
A provider reporting a blended score from unknown prompts, run an unknown number of times on unknown dates, leaves too much hidden. Improvement might reflect changed answers. It might reflect a changed test. The dashboard alone does not settle it.
A credible baseline should disclose the fixed query list, engine panel, run count per query per engine, dates and locations. It should also distinguish mentions, citations, links and accuracy rather than compressing them into one cheerful line.
I cover that disclosure in more detail in why AI visibility reports disagree and what your provider should report. For this survey, the takeaway is simpler: ask who chose the prompts, what buying decision each represents and what changed since last month besides the score.
Turn the findings into four shortlist checks
The teardown leaves me with a useful screening script, not a league table of agencies. If I am choosing ongoing SEO and AI search support, I want the reporting weaknesses exposed here addressed before the contract starts.
1. Separate rankings from AI visibility
Require distinct evidence for Google rankings and AI visibility. They are not interchangeable measures, and a blended score conceals the information needed to judge either one.
I want to see which measure changed and the evidence behind it. “Search visibility improved” is a heading, not an explanation. I keep a ranked provider shortlist in Best Ongoing SEO and GEO Services: Ranked by Who Does the Work, but this reporting test comes before any ranking.
2. Demand the material behind the score
Ask for prompts, raw dated answers and engine context. Add run counts and the third-party sources the provider actually worked on.
I would walk away from a pitch that guarantees a specific spot in ChatGPT or AI Overviews, or supplies a visibility score without prompts, answers or sources. A polished chart is not a substitute for an inspectable test.
3. Connect the work to the tracked questions
Make the provider map its work to buying questions. Content, technical fixes and authority building should connect to prompts that matter to the business, not just to a generic visibility line going up.
This is where the work log earns its place. What changed? Which question was that change intended to address? What did the next set of observations show? I want the reasoning, not a list of completed tasks dressed as a commercial result.
4. Agree on attribution limits before the pitch becomes a promise
Require an explicit account of what the provider cannot prove. Define which sessions are identifiable, how buyer self-report reaches the CRM and what connection to pipeline remains uncertain.
Do that before a monthly report turns an observed mention into “influenced revenue” by enthusiasm alone. Hire the provider that can explain the missing links without pretending they are either proof of success or proof of failure.
Apply the same scrutiny to groas
I work with groas, so run the same test on us. groas reports Google visibility and AI visibility weekly, with the work log behind it. That is a scope commitment, not proof that the next mention adds a deal. It tells you what ran, what changed and what was observed.
The survey’s 77% does not prove revenue, and neither does our reporting cadence. No outside party controls what a model cites. A provider guaranteeing citations, rankings, traffic or leads from ChatGPT or AI Overviews is selling certainty it cannot own. I would walk away from that pitch even if it came from my own side.
Before hiring groas, ask for a sample reporting structure with prompts, engines, dates and run counts attached. Agree on the commercial measures in advance: which AI-referred sessions we track, how self-reported AI discovery reaches the CRM and what stays unmeasured.
Then examine whether recommendations on buying questions move, alongside the commercial records. Do not let a blended score stand in for either. That is the version of this service I trust enough to put my name on.
Overall verdict: keep the warning, keep the denominators
The Reddit summary is worth saving because it makes two weaknesses visible: citation reporting is more common than pipeline reporting among the 22 agencies answering that question, and 11 of 20 third-party tool users do not fully understand prompt selection.
That is a useful warning, not market-wide proof. The report starts with 47 respondents at 46 agencies, mostly small and UK-based, then narrows to different subgroups for different questions. Dropping those denominators makes the percentages sound broader than they are. Turning them into revenue claims goes further still.
I would keep this survey as a list of questions for the next provider. Anyone quoting 77% or 32% as what the industry delivers is reading a flashlight beam as daylight.
What I keep from it fits on one page taped to my monitor:
- Show recommended versus mentioned versus linked, with the prompt and full answer attached.
- Show rankings and AI visibility separately, with dates, engines and run counts where relevant.
- Show identifiable AI sessions, self-reported discovery and CRM pipeline as separate records. State what still cannot be connected.
Groas fits that page better than the opaque reporting practices exposed here: changes come with their reasoning in plain language, and a senior person owns the result. That is the accountability described in the groas philosophy, and it is something a buyer can check rather than take on trust.
I still would not buy pipeline promises from us or anyone else. Buy the disclosure, then pay for the work that moves buying questions.

