Back to blog Several people working with laptops at a shared table

GEO Analytics

Personalization and memory: comparing AI visibility across profiles

2026-10-10·8 min read

A recommendation obtained from the analyst's account describes an answer under that account's conditions. It does not establish what every buyer will see. If the account carries preferences, instructions or information from earlier conversations, copying the same prompt into another account does not guarantee the same starting point either.

For buying-research and GEO analytics teams, the question is specific: does the observed recommendation change when a documented personalization condition changes, while the rest remains comparable? The useful deliverable is a paired comparison, with evidence about the profiles included and the factors that could not be isolated.

This guide studies persistent context between conversations. Measuring recommendations across conversation turns follows a different unit: a trajectory within one dialogue. Neither design turns a favorable screenshot into an estimate for the entire market.

Distinguish profile, account and run condition

A profile defines needs and constraints relevant to the study. An account is the environment used to run it. A condition describes which personalization is available in that environment during a run. Do not use these three terms interchangeably.

You can define one synthetic profile around ease of implementation and another around integration requirements. Those definitions help contrast criteria; they do not establish that two buyer segments exist or have a particular size. For participant profiles, agree which information may be collected and keep personal identifiers separate from analytical answers.

Record which part of a profile appears in the question, which is supplied through instructions and which remains in a persistent mechanism. A need stated explicitly in the prompt is visible context, even if the account also remembers it. That duplicate exposure prevents a clean memory contrast.

Verify available controls, not the mode's name

ChatGPT's memory documentation, checked on October 8, 2026, explains that available controls and personalization sources vary. Remembered information is not a complete transcript and can change. A settings screenshot therefore does not describe every piece of context used.

The current temporary-chat guide distinguishes Personalized and Unpersonalized. The former can use existing personalization; the latter does not use memory, custom instructions or plugins. Neither creates or updates memory while the chat remains temporary. Do not label every temporary chat as an unpersonalized control.

These distinctions call for a configuration record, not a promise of perfect isolation. Check which options exist in the test account, what was selected and which organizational restrictions apply. If the interface cannot isolate memory from instructions or other features, call the contrast a configuration change, not an exclusive memory effect.

Do not transfer one interface's controls to another provider or an API. An API evaluation with explicitly supplied context is a different environment; it does not establish the behavior of a consumer account with persistent memory.

Build a panel you can describe without exposing people

Define the panel before running it: needs, inclusion criteria, the origin of constraints and the conditions you will compare. For a synthetic panel, retain the complete profile text and its version. For participants, collect only what the agreed objective needs and document permission and limits on use.

Do not copy private histories, customer emails or files to make a profile seem more realistic. An approved minimal summary may be enough. Exclude sensitive data that is not essential, and do not infer hidden preferences from an assistant's answer.

Use authorized test environments and follow service terms; this protocol does not require fictitious identities or shared credentials. If existing accounts are your only option, state that you are comparing those accounts and their known configurations. Uncontrolled differences are a design limitation.

Keep market and language fixed within a contrast. For multiple markets, design country-language cells first. Changing location and memory together does not let you attribute a difference to either one alone.

Pair conditions and declare what changes

A pair identifier should connect profile, intent, question text, surface, execution window and repetition. Each side receives its planned condition; do not combine similar questions or runs separated by an update without recording the change.

Proposed contrast	What stays fixed	What you can interpret
Same profile, memory available versus unavailable	Question, instructions, mode and other factors the interface can hold fixed	Association with the documented memory change, if no other feature changed
Personalized versus unpersonalized temporary chat	Question, surface and execution window	Difference between complete configurations, not memory alone
Two profiles under the same configuration	Intent, surface and analytical rules	Difference between the defined profiles, without estimating real segments

Verify conditions before each block and alternate their order when the environment permits it. Always running the control first can confound order with condition. If using the account changes it, neither alternation nor a new chat guarantees restoration of its initial state.

A process check can ask about a noncommercial synthetic detail specified in the profile. It helps identify an unexpected configuration. Recalling the detail does not prove which mechanism supplied it; omitting it does not prove it was unavailable. Keep this check alongside settings evidence rather than treating it as an infallible memory detector.

Treat carryover as a process finding

A run may alter the environment later questions will use. Separate what you know became available from what you cannot inspect. Record observable state before and after a block, along with any additional operator interaction.

Observed signal	Treatment in the record
A brand preference introduced before measuring discovery	Mark prior exposure; analyze outside the unprompted block
A setting or feature changes during the pair	Invalidate the planned comparison; retain both answers and the reason
One side fails to return an answer	Incomplete pair; do not count failure as brand absence
Persistent state cannot be reproduced	Declare an account/configuration contrast, not an isolated-variable contrast

Do not ask a participant to erase their history to fit the study. Any cleanup in a test environment must be authorized and recorded; deleting material is not automatic proof that every prior influence has gone.

If you cannot control carryover, retain run order and narrow the conclusion. You may study a panel under natural use, but that is a different design from a reproducible intervention. Do not fill the record with a clean state you did not verify.

Calculate differences within evaluable pairs

Set an explicit-recommendation rule and apply it to both answers. Keep mentions without recommendations, explicit exclusions and unevaluable cases separate. Preserve the passage supporting each label; a formatting change should not silently alter the rule.

For a brand, call the compared conditions A and B. Among pairs with two evaluable answers, count recommended only in A, only in B, in both and in neither. The paired difference in proportions is:

(B-only pairs - A-only pairs) / complete evaluable pairs

Publish counts for all four cells, not just the subtraction's result. Concordant pairs also belong in the denominator. With no complete pairs, the difference is undefined. Add coverage as complete pairs divided by planned pairs, with reasons for losses.

Calculate by profile and condition first; then decide whether a summary is useful. A profile run more often should not receive more weight by accident. For an aggregate, fix the weights and show individual views. Repetitions of the same profile share context: intervals and tests must respect that dependence rather than treating each answer as another person. The answer-variability protocol addresses noise within a configuration, not panel representativeness.

Separate brand familiarity from discovery

A profile containing the target brand already exposes the system to it. Even when the current question does not name it, the result is not discovery without prior exposure. Keep one block without commercial names and another with known preferences; do not combine their rates as though they measured the same outcome.

In the exposed block, inspect how the answer handles that preference: retaining it, suggesting alternatives, excluding it or merely repeating it. Repeating the name is not enough to label a recommendation. Also review whether the reasons meet the defined requirements and describe the offer correctly.

This separation helps decide whether to investigate suitability, prior preference or an incorrect description. It does not justify manufacturing memories to favor the brand or establish that a GEO intervention caused the observed change.

Report the panel, not an invented audience

The report should reconstruct each pair: profile and version, condition, observable configuration, question, order, date, answer, label, passage and exclusion. Retain full synthetic-profile text and an approved minimal participant record; restrict access to identifiable material.

Show differences by profile, complete and incomplete pairs, prior brand exposure and controls you could not hold fixed. State which decision the evidence supports: extending a profile, reviewing a product claim or repeating a contrast with better controls. A rate from your panel does not indicate what share of all buyers will see the brand.

If you use Mentio to observe visibility, keep this record alongside the measurement and check which configuration and evidence each analysis provides. This guide does not attribute account cloning, private-memory access or automatic buyer simulation to the product. Profile comparisons require a design you can explain and review, even when parts of the execution are automated.

Frequently asked questions

Does starting a new conversation remove personalization?

It does not establish that. Document dialogue history separately from persistent controls available in the account. If you cannot verify the state used, do not describe the run as an unpersonalized environment.

Is temporary chat a control without memory?

Not always. Current ChatGPT documentation distinguishes personalized and unpersonalized temporary chat. Record which option was selected, and do not attribute a contrast that changes other features to memory alone.

What can synthetic profiles establish?

They can describe outcomes under the needs and conditions you defined. They do not estimate real segment sizes or every buyer's behavior without a supporting design and sample.

What should I do when one answer in a pair fails?

Keep the attempt and reason, and mark the pair incomplete. Do not include it in the complete-pair difference or count the failure as a valid answer without a recommendation.

Can I combine profiles that already know the brand with those that do not?

Not in a rate presented as discovery without prior exposure. Separate the blocks, record where the brand was introduced and compare each under its own conditions and denominators.

Want to know if AI mentions your brand?

Discover your visibility in ChatGPT, Claude and Gemini in minutes.