Back to blogAnalytics dashboard on a laptop used to monitor an AI product launchGEO Analytics

    How to Monitor a Product Launch in ChatGPT, Gemini and Perplexity

    2026-07-29·13 min read

    On launch day, your website, announcement, partner pages, comparisons and market conversations begin telling a new story. ChatGPT, Gemini and Perplexity may pick up parts of that story at different speeds. They may also retain the old name, mix two versions, or claim that the product is not yet available.

    One isolated check cannot tell you whether the launch is being understood. You need to compare what models said before publication, what changes during launch and what becomes stable afterward.

    This guide turns monitoring into a time-bound operation. It starts with an AI visibility tracker that is already configured and adds what a launch requires: a dated fact sheet, risk prompts, a more intensive cadence, alerts and owners.

    Why a launch is not ordinary tracking

    Steady-state measurement looks for weekly or monthly trends. A launch is a known event that changes many sources at once and concentrates risk into a few days.

    The team needs to answer four questions:

    1. Discovery: Does the product appear when someone asks about the problem or category?
    2. Understanding: Does AI explain what it is, who it serves and how it differs?
    3. Freshness: Does it recognize current availability, name, version and facts?
    4. Risk: Does it invent, merge or retain facts that could damage a decision?

    You are not demanding that models update instantly. You are observing the lag, locating where the story fails and deciding which evidence or source needs attention.

    Define launch truth before opening the dashboard

    Do not score answers against what each person remembers from the campaign. Create a launch truth sheet approved by product, marketing and, when necessary, legal.

    FieldExampleRisk if wrong
    Canonical nameMentio SignalConfusion with the parent brand
    CategoryAI visibility measurement platformIncorrect classification
    AudienceB2B marketing teamsRecommendation to the wrong audience
    ProblemDetect what AI assistants say about a brandDistorted promise
    DifferentiatorRecurring comparison across models and competitorsGeneric message
    AvailabilitySpain and Mexico from July 29Offer shown in an unsupported market
    Price or planOnly if public and stableIncorrect commercial decision
    Previous-product relationshipReplaces X; complements YVersion confusion
    Approved sourcesProduct page, documentation and announcementFragmented evidence
    OwnerProduct MarketingContradictions go unresolved

    The sheet does not replace an AI brand entity audit. It uses that audit as its foundation and limits evaluation to facts that change with this launch.

    Assign each fact:

    • high, medium or low criticality;
    • a primary URL;
    • an effective date;
    • markets and languages;
    • acceptable wording;
    • phrases that indicate an error or confusion.

    Build a launch-specific prompt matrix

    Do not add the product name to twenty prompts and call it monitoring. Combine branded and unbranded questions to observe discovery and understanding.

    BlockWhat it testsExample
    Unbranded problemDiscoveryHow can I learn what AI assistants say about my brand?
    CategoryClassificationWhich platforms measure visibility in ChatGPT, Gemini and Perplexity?
    ComparisonDifferentiationCompare the new product with two real alternatives
    Branded definitionUnderstandingWhat is [product] and who is it for?
    AvailabilityFreshnessIs [product] available in Mexico?
    CompatibilityFitDoes it integrate with [relevant system or workflow]?
    MigrationVersion relationshipShould I keep using [previous product]?
    Price or planCommercial riskHow much does [product] cost?
    Identity confusionDisambiguationDoes [product] belong to [company]?
    ControlEnvironmental stabilityA related question that should not change

    Start from a versioned AI visibility prompt bank, then create a launch cohort. Label intent, market, language, model, criticality and expected fact. Freeze wording and configuration before T-7 so a prompt edit cannot look like adoption.

    Phase 1: record the baseline from T-14 to T-4

    The baseline describes what models knew before public exposure. Run the cohort under comparable conditions and retain full answers.

    Record for every run:

    • date and time;
    • prompt and version;
    • model, mode and web access;
    • country and language;
    • complete answer;
    • brand and product mentioned;
    • position or prominence;
    • cited sources and URLs;
    • narrative classification;
    • correct, missing or wrong facts;
    • auditable capture or reference.

    Repeat every condition several times. Generative systems vary, so apply the AI answer variability protocol within each model and market. A baseline is not one answer collected the night before.

    This phase should separate three states:

    1. The product does not yet appear, as expected.
    2. It appears because public information already exists.
    3. It appears with incorrect or prematurely released facts.

    The third state needs immediate review. Do not wait until T0 to discover an indexed page with the wrong name, price or availability.

    Phase 2: check readiness from T-3 to T-1

    Before launch, validate that approved sources are publishable, consistent and accessible. You are not measuring success yet; you are checking that the evidence system is not broken from the start.

    Review:

    • one canonical URL per product and market;
    • consistent name, category and description;
    • dates and availability without contradictions;
    • aligned documentation and FAQ;
    • valid structured data where appropriate;
    • links from the parent brand and related products;
    • partners or distributors using the correct version;
    • no old page that still looks current.

    Run a short wave at T-1 with the baseline configuration. Save any pre-announcement change. If you change prompts, sources or conditions, create a new version and do not mix its results into the baseline.

    Phase 3: monitor the T0 to T+3 window

    Record the public launch time. From that point, run dated waves rather than improvised checks by different team members.

    A practical cadence:

    TimeObjective
    T0Capture the state at publication
    T+6 hoursDetect early critical errors
    T+1 dayCompare initial adoption by model
    T+3 daysSee whether the signal persists or was a spike

    Not every team needs a run every six hours. Reserve high frequency for critical facts, priority markets or campaigns with substantial exposure. Keep enough repetitions to avoid mistaking one answer for a change.

    Calculate coverage before interpreting each wave:

    coverage = valid runs / planned runs

    If a model or configuration fails, mark the wave incomplete. Do not present an aggregate improvement built from fewer observations.

    Classify adoption, absence and error

    A binary mention does not explain whether the launch was understood. Classify each answer with a shared taxonomy:

    StatusCriterion
    Adopted and correctName, category, audience and main fact match
    Correct but incompleteNo errors, but a relevant element is missing
    StaleKeeps an old version, price, date or availability
    ConfusedMixes the product, company, version or competitor
    IncorrectStates a fact that conflicts with the approved sheet
    Mentioned without contextNames the product but does not explain its fit
    AbsentDoes not appear in a question where it could be relevant
    Not evaluableEmpty answer, blocked run or invalid configuration

    Separate mention rate from narrative adoption rate:

    adoption rate = adopted and correct answers / valid answers

    A brand can gain mentions while accuracy gets worse. Show critical-error volume and coverage next to every rate.

    Measure the delta against baseline

    For every wave and condition, compare:

    • change in mention rate;
    • change in narrative adoption;
    • change in position or prominence;
    • number of correct facts;
    • new and resolved errors;
    • newly cited sources;
    • confusion with previous products;
    • dispersion across repetitions.

    Do not merge models and markets automatically. A launch may be well understood by Perplexity in Spain and poorly described by ChatGPT in Mexico. Preserve the condition-level view and aggregate only with explicit business weights.

    The delta shows a temporal association, not causality. If you need to attribute a change to a specific intervention, design a GEO experiment with controls and an expected mechanism.

    Define alerts before seeing results

    Escalation rules should exist before an uncomfortable answer appears.

    LevelExampleAction
    CriticalFalse price, availability, safety or identityOpen an incident immediately and assign an owner
    HighRepeated confusion in two models or a priority marketReview sources the same day
    MediumIncomplete message repeated across two wavesImprove evidence and observe the next wave
    LowOne absence or position variationRecord without reacting

    One possible rule:

    • Escalate any critical error even if it appears once.
    • Escalate a high-severity error if it repeats in the same condition or appears in two models.
    • Do not escalate one absence without reviewing repetition, relevance and coverage.
    • Close an incident only after the source is corrected and two comparable waves no longer reproduce the problem.

    Correction may require updating a primary page, aligning documentation, removing an old URL or contacting an external source. Use the broader guide to fixing wrong AI information about a company for the general process; here, every action must remain tied to a launch incident.

    Phase 4: confirm stabilization from T+4 to T+14

    Reduce frequency and retain comparable checkpoints at T+7 and T+14. Look for stability, not a perfect score.

    The launch can close when:

    1. Planned coverage is complete.
    2. Primary metrics do not change materially across two waves.
    3. No critical errors remain open.
    4. The correct narrative appears consistently in priority conditions.
    5. Dominant sources have been identified.
    6. Every incident has a resolution, owner or explicit risk acceptance.

    Save the closing wave as the new operating baseline. Retire prompts tied only to a date or temporary availability; category, comparison and product-relationship prompts move into recurring tracking under a new version.

    Example: launching a module for a B2B SaaS platform

    A company launches an analytics module for customers of its main platform.

    At T-7, the new name is absent, but ChatGPT assigns its function to the old product. At T0, Perplexity cites the announcement and describes the function correctly, while Gemini keeps the previous architecture. At T+1, two answers claim that the module is sold separately even though it is available only within a specific plan.

    The team does not conclude that "Gemini is worse" or rewrite the entire campaign. It opens a commercial incident, discovers ambiguous wording on the pricing page, corrects it and aligns the documentation. At T+3, it repeats the same cohort. The error disappears in Perplexity and ChatGPT but remains in one Gemini repetition, so it stays under observation until T+7.

    The record connects a specific answer to a fact, source, owner and next check.

    Assign owners and keep one incident log

    RoleResponsibility
    Product MarketingTruth sheet and narrative
    ProductFeatures, compatibility and versions
    SEO/GEOCohort, runs, sources and links
    PR/CommsAnnouncement and external sources
    Legal/ComplianceRegulated or sensitive facts
    AnalyticsQA, deltas and coverage
    Incident ownerCorrection and closure

    Use one log with:

    incident_id, date, prompt, model, market, answer, affected_fact, severity, likely_source, owner, action, status, next_measurement

    The leadership summary should contain changes, risks and decisions rather than hundreds of screenshots. Move the closeout into an AI visibility executive report and keep the dataset as an auditable appendix.

    Mistakes that invalidate launch monitoring

    1. Starting on launch day without a baseline.
    2. Changing prompts between waves without creating a version.
    3. Running only one answer per condition.
    4. Treating any mention as correct narrative adoption.
    5. Merging countries, languages and models into one percentage.
    6. Ignoring incomplete coverage or access failures.
    7. Reacting to every absence as a crisis.
    8. Failing to separate the new product, old version and parent brand.
    9. Correcting sources without recording the incident that caused the change.
    10. Closing monitoring without turning T+14 into the new baseline.

    Operational checklist

    • [ ] The truth sheet is approved and dated.
    • [ ] Every fact has a source, criticality and owner.
    • [ ] The cohort covers discovery, understanding, freshness and risk.
    • [ ] Prompts and configuration are frozen before T-7.
    • [ ] There is a repeated baseline by model, market and language.
    • [ ] T0, T+1, T+3, T+7 and T+14 are scheduled.
    • [ ] The taxonomy separates adoption, absence, staleness, confusion and error.
    • [ ] Alerts are defined before results are reviewed.
    • [ ] Every incident retains the answer and evidence.
    • [ ] Closeout creates a new baseline and an open-issues list.

    To compare how long each fact takes to be observed, add a content change latency record. Keep the first positive result separate from persistence across later waves.

    FAQ

    When should I start monitoring an AI product launch?

    Start seven to fourteen days before the public date. That window lets you record a baseline, check for premature or incorrect information, and freeze the sample before the campaign, press coverage and product pages change the environment.

    Which prompts should launch monitoring include?

    Include category, problem, comparison, compatibility, availability, price, migration and previous-product relationship prompts. Add controls that should not change and risk prompts about facts that would cause harm if AI explained them incorrectly.

    How often should I run prompts during the launch?

    Use a more intensive cadence from T0 to T+3, with multiple repetitions per condition, then reduce frequency as answers stabilize. Keep comparable checkpoints at T-7, T-1, T0, T+1, T+3, T+7 and T+14.

    How do I know whether AI has adopted the launch narrative?

    Adoption requires more than mentioning the name. The answer should correctly connect the product, category, audience, availability and main proposition without confusing it with a previous version. Measure that match against an approved fact sheet.

    Should one incorrect answer trigger an alert?

    Record every incorrect answer, but not every one requires an incident. Escalate critical price, availability, safety or identity errors immediately; for lower-severity errors, require repetition in the same condition or occurrence in more than one model.

    When can the launch be considered stable?

    When planned coverage is complete, primary signals stop changing materially between waves, no critical errors remain open, and the fact sheet appears consistently in priority conditions. Document closure and keep the dataset as the new baseline.

    Outside the launch timebox, hand monitoring over to the recurring operation described in AI visibility drop alerts.

    A launch is a window with a known date. For steady-state catalog tracking outside that window, use the protocol in tracking AI product recommendations at SKU level.

    Monitor the launch with Mentio

    Monitor your AI visibility with Mentio ->

    Want to know if AI mentions your brand?

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

    Related articles