Do Persona Prompts Change How AI Searches? Evidence from 1,937 Grounded Search Queries

At a glance

Research snapshot

Research question

Why this question matters

Grounded AI systems do more than generate prose.

When search is enabled, the model may formulate one or more queries before or during answer construction.

That creates an observable layer between:

prompt

and:

final answer.

For AI visibility, AEO and GEO research, this layer matters because two prompts that lead to different search formulations may expose different information needs—even when they originate from the same underlying business question.

But query telemetry is easy to overinterpret.

A different query does not automatically mean:

• A different publisher was retrieved
• A different source influenced the answer
• The model performed deeper research
• The query was better
• The final answer was more accurate

Our earlier research on Gemini query fan-out asked a different question:

When similar prompts are phrased differently, does Gemini search for similar information?

That study found a strong relationship between prompt similarity and query-set similarity.

This study holds the underlying scenario constant and changes the persona condition.

The question is narrower:

Does the decision context supplied by the persona correspond to a different recorded search formulation?

Study design

Study design

The parent experiment began with 150 neutral business scenarios.

Each scenario was rendered under five persona-conditioning bundles:

• CMO
• VP of Product Marketing
• SEO Manager
• Demand Generation Director
• Founder/CEO

Gemini 3.6 Flash answered with Google Search grounding enabled.

A later quality audit excluded one malformed response, leaving 749 substantive responses.

Primary matched query comparisons used the 149 complete scenarios with all five persona conditions present.

Each complete scenario creates ten unordered persona pairs, giving:

149 × 10 = 1,490 primary pair comparisons.

Across the substantive responses, the grounded output contained 1,937 usable search-query entries.

All 1,937 passed the query-quality screen.

There were:

• No within-response duplicate query entries
• Only nine normalized queries shared across persona conditions within matched scenarios

The analysis compared query sets at several levels because exact-string comparison alone can create a misleading ceiling.

Finding 1

The same scenario almost never produced the same literal search queries across personas

The first comparison used normalized exact query sets.

The metric is:

1 − query-set Jaccard similarity

where:

• 0 means identical query sets
• 1 means no exact normalized query overlap

Mean divergence was:

0.9991

with a 95% scenario-bootstrap interval of:

0.9978–0.9999.

That means exact normalized search-query overlap across persona conditions was almost nonexistent.

But this result needs careful interpretation.

Exact query strings are brittle.

Two queries can use different strings while expressing closely related information needs.

So exact divergence establishes that literal queries were different.

It does not tell us how different their content was.

Finding 2

The query differences remained substantial after moving beyond exact strings

To measure degree rather than literal identity, we used two graded representations.

Token-set divergence

Mean divergence:

0.6822

95% CI:

0.6710–0.6933

TF-IDF lexical divergence

Mean divergence:

0.8870

95% CI:

0.8798–0.8938

This matters because the result is not simply:

“The strings were different.”

The recorded queries also differed substantially in their vocabulary and formulation.

Open full-resolution figure

Finding 3

Different persona conditions produced different search-language signatures

The persona differences visible in final answers were also visible upstream in the query language.

Using a fixed vocabulary codebook defined before the retrieval outcomes were inspected, we compared technical, commercial and governance-oriented language in the recorded queries.

SEO Manager queries were much more technical

SEO Manager recorded roughly 74.9 technical terms per 1,000 query words, compared with single-digit rates for the other four conditions.

Demand Generation queries were much more commercial

Demand Generation recorded approximately 75.6 commercial terms per 1,000 query words.

CMO queries were much more governance-oriented

CMO recorded approximately 35.9 governance terms per 1,000 query words.

These are not measurements of how real SEO managers, demand-generation leaders or CMOs search.

They are properties of the model's recorded query formulations under the historical persona-conditioned prompt bundles.

Pair distances

Which persona pairs searched most differently?

Using lexical divergence:

The strongest observed lexical query divergence was:

Demand Generation Director / SEO Manager

The weakest was:

CMO / Founder/CEO

That mirrors the broader persona programme.

SEO Manager tended to occupy the most distinctive region of the response and query space.

CMO and Founder/CEO were repeatedly among the closest conditions.

Again, “closest” and “farthest” describe representational distance—not quality or real-role similarity.

Finding 4

More different search language was associated with more different answers

The parent analysis had already created a persona-residual answer distance after masking direct persona labels and profile phrases.

We joined those answer distances to the query divergences for the same:

scenario × persona pair

This let us ask:

When two persona conditions formulate more different searches, do they also tend to produce more different final answers?

Exact query divergence was almost saturated.

Its within-scenario association with residual answer distance was:

r = −0.012

95% CI:

−0.043 to 0.016

That is essentially no association.

The graded query measures were different.

Token divergence ↔ residual answer distance

r = 0.514

95% CI:

0.459–0.566

Lexical query divergence ↔ residual answer distance

r = 0.495

95% CI:

0.436–0.553

The relationship also appeared descriptively across the ten persona-pair profiles.

Pair-profile Spearman correlations were:

• Token divergence: 0.939
• Lexical divergence: 0.818

In other words, persona pairs that were more differentiated in query language also tended to be more differentiated in their final residual answers.

This is an association.

It is not evidence that the query difference caused the answer difference.

Search and answer generation occurred inside the same grounded call.

Open full-resolution figure

Topic stability

The pattern remained high across all five topic clusters

Lexical query divergence remained high across all five clusters:

Range:

0.864–0.909

The exact magnitude varied, but no topic cluster showed a collapse in persona-associated query differentiation.

This does not make the effect causal.

It does show that the observed query differentiation was not confined to one subject area.

AEO research

What this means for AI visibility and AEO research

The practical implication is not:

“Track every query Gemini generates.”

The more useful conclusion is:

The query layer can reveal variation that the final answer alone may hide.

If two persona conditions ask different grounded search questions before producing an answer, then monitoring only the final response can miss an important part of the observable process.

For research teams, this suggests a layered evaluation model:

1. Prompt layer

   What did we ask?

2. Query layer

   What information needs did the grounded system formulate?

3. Source layer

   Which evidence did it retrieve or cite?

4. Answer layer

   What did it finally say?

The current study observes layers 1, 2 and 4.

Its source layer was not resolvable.

That distinction should remain explicit in any AEO or GEO workflow.

Panel design

Implications for persona-panel design

The query findings also help explain why persona panels can remain diverse even when the underlying scenario is held constant.

The five persona conditions did not simply apply different wording to the same apparent search plan.

Their recorded search formulations themselves differed.

That means persona-panel diversity can arise before answer synthesis.

But this still does not establish that every query difference is useful.

A more technical query may be more relevant, less relevant, redundant, overly narrow or simply different.

The next step for stronger research would be to connect persona-conditioned query formulation to reliably resolved source pools, claim-level evidence, factual accuracy and recommendation quality.

Those links are not identified here.

Conclusion

Conclusion

Persona prompts changed more than the final wording in this historical grounded Gemini experiment.

Across 149 matched scenarios, different persona conditions were associated with substantially different recorded search-query formulations.

The difference was visible at three levels:

• Almost no exact normalized query overlap
• Substantial token divergence
• High lexical divergence

The query language also reflected distinct persona-associated emphases.

SEO Manager queries were more technical.

Demand Generation queries were more commercial.

CMO queries were more governance-oriented.

And persona pairs with more differentiated query language tended to produce more differentiated residual answers.

But the research stops at the boundary the data supports.

We cannot say the different queries caused the different answers.

We cannot say different publishers were retrieved.

And we cannot say more differentiated searches were better searches.

The strongest conclusion is:

In this matched grounded-AI run, persona-associated differentiation was visible in the system's recorded search-query formulation before it was visible in the final answer.

For the broader programme—including prompt-echo robustness, task persistence and panel-size analysis—see Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses.

Research context

Related Kojable research

• Persona Prompts Change More Than Tone: Evidence from 750 Grounded AI Responses

  The V2 capstone.

• How Many AI Personas Do You Need? What a Five-Persona Panel Actually Preserves

  The panel-size companion.

• How Predictable Are Gemini's Fan-Out Queries? A 180-Prompt Study

  Prompt similarity and query-set similarity.

• What Query Fan-Out Can—and Cannot—Tell GEO Teams

  Interpretation boundaries for query telemetry.

• Beyond Persona Labels: Do AI Personas Still Differ After Explicit Cues Are Removed?

  Cue-ablation persona research.

• AI Answer and Source Research

  Browse the Kojable Research index.