The Gmail Test With 1,922 Responses Has a Math Problem. The Strategy Still Matters

When someone connects Gmail to Google AI Mode, what they have seen in their inbox can change the brands Google recommends later. The effect appears in an iPullRank experiment, but it is not yet a universal ranking signal. The test used lab accounts, eight categories, and Personal Intelligence enabled. It does not represent everyone using Google Search.

The strategic conclusion still matters: lifecycle emails are part of the context a brand can create. The numerical conclusion needs more care.

What the experiment actually found

Google lets a person connect Gmail and Google Photos to AI Mode through Personal Intelligence. The connection is optional, off by default, and Google says the model does not train directly on the inbox or photo library. Responses can use that information to tailor recommendations to the person's own context.

The iPullRank team wanted to know whether a brand mentioned in Gmail would be more likely to appear in AI Mode recommendations. It set up three accounts: a blank account without Personal Intelligence, a blank account with Personal Intelligence, and a mature personal account with a long history. The team then seeded test brands through email and photos, ran recommendation prompts across eight categories, and extracted brand mentions, positions, and citation URLs from the responses.

The study collected 1,922 AI Mode responses and 22,064 brand-level rows between March 30 and April 15, 2026. The direction was clear:

• seeded brands appeared in 66.8% of relevant responses in the Personal Intelligence account;
• the control account showed them in 23.9%;
• brands seeded through email appeared in 53.6% of responses;
• brands seeded through photos appeared in 10.5%.

That matters. The personal-context layer changed recommendations, and Gmail was a stronger signal than photos in this test.

The 46-point headline does not add up

The iPullRank summary says seeded brands were 46 percentage points more likely to appear than in the control condition. The figure appears to come from its difference-in-differences method, not from subtracting the before-and-after appearance rates directly. The article does not show the full formula or enough denominators to reconstruct that value from the summary alone.

That does not make the finding false. It changes how the result should be reported. I would use the visible rates, 23.9%, 66.8%, 53.6%, and 10.5%, and label the 46-point result as the authors' model output rather than a simple arithmetic comparison. I would treat the top-three and top-ten changes as study metrics, not as deltas derived from the headline rates.

The gap between 46 and 42.9 matters because a growth promise needs a reproducible calculation. The direction of the finding is useful. The magnitude needs a methodological clarification.

Why email can change a recommendation

An AI Mode response is not built only from public web pages. The model can combine web signals with information a user chooses to share from their own apps. In a question about clothing, for example, a purchase, booking, or comparison email can help identify the category, preferred brand, and use case.

That turns certain emails into context material. They are not just campaign messages. They are small records of a commercial relationship:

• a welcome email explains which category the brand belongs to;
• an educational sequence teaches the problem the brand solves;
• a receipt or renewal connects the brand to a real purchase;
• support leaves instructions the person may need again;
• a comparison email explains a product difference clearly.

The iPullRank hypothesis makes sense for B2B and services. Someone who has seen several emails from a company may recognize that brand when AI Mode answers a category question. A new brand without a verifiable public footprint does not get the same advantage just because someone emails it. In the study, real brands already had products, reviews, comparison pages, and other public signals. Fictional names appeared in fewer cases and did not establish a complete public entity.

What I would do with this lifecycle signal

I would not send more email. That is the wrong conclusion. I would treat each email as a piece of education worth saving, forwarding, or revisiting later.

Start by mapping the messages we already have. For each flow, ask three questions:

1. What category and use case can someone understand from this email?
2. What proof makes the brand credible, rather than merely named?
3. Is the information still useful when someone opens it three months later?

Then prioritize four surfaces.

Onboarding. The first message should explain what problem the product solves and who it is for. It does not need to repeat the entire pricing page. One concrete example makes the association clearer.

Education. A useful sequence teaches a method, answers an objection, or shows how to make a decision. That explanation may be what someone needs again when AI Mode answers the same question later.

Post-purchase and support. Confirmations, renewals, requirements, and support replies create a record that belongs to the relationship. They should be accurate and easy to find, not stuffed with keywords.

Comparison. An email that explains when an option fits and when it does not can help a customer choose with more confidence. It also reduces the risk that AI Mode describes the brand in generic or incomplete terms.

After publishing or changing an email, I would not check one prompt and declare victory. I would repeat the same panel across several platforms, save the full answers, and review the public footprint. The method in A ChatGPT citation is a sample, not a result applies here.

The risk of turning a signal into spam

If every brand starts filling inboxes with recommendation messages, the effect may fade or become unpleasant. Google presents Personal Intelligence as personalization, not an advertising system based on a buyer's inbox history.

Sending an indiscriminate campaign to “activate” a brand does not solve the problem. It can:

• increase complaints and unsubscribes;
• reduce trust with the person receiving the message;
• make the email irrelevant to the query;
• create an association AI Mode cannot validate;
• ignore the user's privacy preferences.

The metric I care about is not how many emails were sent. It is the quality of the context a person keeps. If a customer cannot explain why the brand is useful to them, the message is not doing its job.

A test brands can defend

If Mintec wanted to apply the finding, I would run it as controlled research, not promise a ranking lift.

First, document twenty real decision questions: price, implementation, integrations, alternatives, use cases, and risks. Second, separate public AI Mode responses from responses run with Personal Intelligence enabled on a test account. Third, use only our own data or information voluntarily supplied by participants. Never use a customer's inbox or private information.

The report should have four columns: prompt, response, cited brands, and exact URL. Repeat the sample on two dates, then add sessions from Google AI, referrals from ChatGPT or Perplexity, and branded searches. Do not present an increase in citations as an increase in revenue unless the business data supports it.

Keep two things separate. The study shows an association under a specific condition. It does not prove that Gmail is a universal ranking factor, that every category responds the same way, or that a company can buy the signal.

The final read

The Gmail finding does not turn email into a GEO trick. It changes the scope of GEO. A brand no longer competes only for visibility on the public web. It also leaves signals in its customers' experience, and those signals can return in a personalized answer.

The useful part is fairly simple: write emails that teach, confirm, and help. The part that requires caution is not turning that work into a ranking promise the data has not earned.

After an experiment of 1,922 responses, the sensible move is to keep the signal and question the headline. The strategy may be good. The math needs to be clear before anyone builds a forecast on it.