A defensive denial, an invite Google won’t accept, and what happens when four AI models are asked to do a spokesperson’s job
By Greg Jarboe
A Google spokesperson told The New York Times that Kevin Indig’s AI Mode research was “flawed” because it relied on a small sample and gave users specific tasks. When a platform dismisses inconvenient research with a one-line technical objection instead of publishing its own numbers, that is not a rebuttal. It is a tell. And it tells SEOs something specific about where citation share is actually being decided right now.
Indig’s team ran a study on AI Mode user behavior. Thirty seven participants completed seven standardized shopping tasks, the kind of research, credit card comparison, and product lookup queries that make up real search volume. Across those sessions, 77.6 percent ended without the user ever clicking through to an external site.
When Kate Conger reported on the finding for The New York Times on July 20, in a story about Google’s shrinking commitment to the open web, Google’s spokesperson pushed back, saying the study did not reflect normal user behavior. Indig responded quickly in a LinkedIn post. He agreed the sample was small. He called the finding directional rather than a census. Then he invited Google to run a larger version of the same study together, since Google is the only party with the query volume to settle the question at scale.
That invitation has not been accepted, at least not publicly, and I don’t expect it will be.
I ran the story by Katie Paine, and she called Google’s explanation palatable, if not great, which is more credit than I was ready to give it. But her real point cut deeper than either side’s. Whether the true figure turns out to be 76 percent or 66 percent, she said, it still points to a massive shift in behavior, and a sample size in the mid thirties is genuinely risky to generalize from. A 74 year old asking what lumber will hold up on a saltwater dock does not search or click the way a 20 something checking an exam answer does, and neither behaves like a 40 something researching the best breed of puppy. Average all three into a single retention number and you lose the thing that actually matters to a marketer, which audience is leaving the search results page and why.
My view is that Google’s spokesperson did more to confirm Indig’s finding than the study itself did. A confident rebuttal comes with better data. A defensive rebuttal comes with a methodology complaint and nothing else. PR people have seen this pattern before, back in the “Before Times” (i.e., before Google), when corporate communications teams learned
that picking a fight with anyone who “buys ink by the barrel” rarely ended the way they hoped.
I wanted to test whether that instinct held up, so I ran a different experiment using the same NYT story as a prompt. I asked a simple question. Could AI do the job of a Google spokesperson. I pulled the AI Overview answer, the AI Mode answer, both variants Gemini generated, and fed all four into Gemini Notebook’s Strategic Framework report.
The result across all four outputs was consistent, and each one got there a slightly different way. The AI Overview answer split the job into what AI can do, draft statements and pull past company data, and what it cannot, handle a crisis with real empathy or carry legal and moral responsibility for a corporate misstep. Both Gemini answers, generated independently, landed on the same verdict using almost the same language, calling AI a co-pilot rather than a replacement. One of them added a detail worth sitting with. Sundar Pichai himself said in late 2025 that a CEO’s job, built on data assessment and autonomous decisions, might actually be easier for AI to handle one day than a spokesperson’s job is. If the person running the company is more replaceable than the person explaining the company, that says something about what the market for AI at the podium actually values.
The Strategic Framework report synthesized all three into a 70/30 model. AI owns the 70 percent, the data retrieval, the localized drafts, the sentiment testing before a statement goes out. A human owns the 30 percent that cannot be automated away, and the report was blunt about why. An AI generated apology reads as hollow and deflects the accountability a crisis actually demands. A hallucinated statistic in a sensitive matter, like the antitrust and content extraction disputes Google is currently navigating, risks a stock sell off or a legal liability that a drafting tool has no way to weigh. The framework’s own
phrase for the arrangement stuck with me, AI as the voice in the ear, human as the one still standing at the podium.
Paine, who was my predecessor in corporate communications at Lotus Development four decades ago, agreed with the AI output on this point too, though for a more specific reason than the models gave. She loves the premise of an AI spokesperson, but told me credibility depends on an authentic voice that can respond to unexpected questions and circumstances at the drop of a hat, with no prompt to fall back on. She was not convinced AI can do that yet, and neither was any of the four AI tools I tested.
Put those two threads together and you get the actual story. Google’s spokesperson gave a short, defensive, single sourced statement to deny a finding it had every ability to disprove with real numbers. Every AI model I tested agrees that this is precisely the kind of moment where a low effort automated response fails and a human with named accountability is what earns trust. Google did not use AI to answer Conger. A person did. And that person’s answer still reads like something optimized for legal safety rather than credibility.
Here is how to apply this to your own strategy.
First, treat single line platform dismissals of inconvenient data as a signal to verify, not a reason to stand down. Pull the primary source yourself before deciding whose numbers to trust, the same way you would build out a Citation Share of Voice analysis before claiming a competitor is winning AI visibility.
Second, run the “could AI do this job” test on the roles inside your own funnel, not just on spokespeople. Ask where your content is already being out cited by AI Overviews and AI Mode, and separate that from the parts of your process, sales conversations, dispute resolution, expert commentary, where a named human is still the reason someone trusts your brand enough to click through or convert.
Third, build your own content around named human sources and verifiable data rather than aggregated claims, in line with AMEC’s GEO Principles. Indig’s study earned coverage in the Times because a real person’s name and a documented method were attached to the number. That is also exactly what earns a citation inside AI Mode today.
My take is that the SEOs who wait for Google to confirm an inconvenient study before acting on it will always be a news cycle behind the ones who verify it themselves and move first. Indig already offered Google a joint study. Until that happens, the smartest move is totreat his numbers as the best available evidence and build your GEO strategy around them
