Back to blogProfessional video camera on a tripod in a studio beside a monitorGEO Analytics

    Product videos cited by AI: evidence and timestamps

    2026-10-28·7 min read

    A citation that opens a product video does not necessarily locate the evidence. An answer may describe a feature that is only spoken, a variant that is only visible or a condition stated shortly before the linked moment. Reviewing the transcript alone leaves part of that evidence out.

    This guide is for owners of demonstrations, documentation and GEO measurement. The deliverable is a record of the video version, tracks examined and verifiable intervals. It supports checking a particular answer, not promising that assistants will cite the video or explaining how to rank it on YouTube.

    The citation support audit establishes how to connect a claim to a passage. Video adds an earlier task: ensuring that visuals, audio and text refer to the same content on the same timeline.

    Fix the edit and its timeline

    Retain the original answer, citation marker, linked URL and final destination. Distinguish the watch page, player and media file. Google documents these as three different resources; opening the page does not establish that the file was inspected.

    Record the title, owner, observed duration, access date and declared version. If you have an authorized copy, identify the file and its hash. The hash helps detect changed bytes; it does not prove when the video was published or which edit the assistant could access.

    Use elapsed time from the start of that edit as the reference and preserve the format of the original locators. Do not confuse player seconds, frame numbers and production timecode. Document the time base and method for any conversion; do not calculate an exact frame from an assumed frame rate.

    If only the current version is accessible, limit the verdict to that edit. An unchanged URL may lead to edited content. Do not reconstruct the version available when the answer was generated with certainty you do not have.

    Identify the tracks under review

    Record which audio, captions and on-screen text were available. Keep the language, origin and version of each track: original speech, dubbing, owner-provided captions, automatic captions or translation. They are not interchangeable copies.

    YouTube warns that automatic captions may misrepresent speech and recommends reviewing them. When a product condition depends on a word, check the authorized audio and retain any discrepancy. Correcting a transcript does not change what is audible in the recording.

    Separate text within the video from player captions and the page description. They may contain different information. Identify the channel for each item of evidence so a condition stated only in the description is not attributed to the demonstration.

    When a required track is unavailable, leave the issue pending and state why. Missing audio, unloaded captions or illegible visuals do not establish that the information is absent from the video.

    Align derivatives with the original

    For an excerpt, retain its in and out points within the reference edit. Also record cuts, speed changes and audio or caption offsets. A derivative needs its own identity and an explicit link to the original.

    For a continuous excerpt with no speed changes or internal cuts, original time is the excerpt's elapsed time plus its starting point in the original. Outside those conditions, retain a segment-level map; a single offset no longer describes the transformation.

    Check alignment against observable visual and audio cues. Do not adjust a track merely until it matches the answer you want to validate. If synchronization cannot be resolved, retain the uncertainty interval and avoid claiming an exact match.

    A still capture must identify its edit and position. By itself, it does not establish a sequence, transition or order of operations. Preserve a replayable interval when the claim depends on movement or a complete action.

    Locate an interval that preserves the conditions

    Keep the citation's suggested time separate from the interval found by the reviewer. Open the linked point and check where playback actually begins. A link near a passage does not guarantee coverage of the complete claim.

    Extend the interval enough to preserve the subject, condition and result. Do not apply a fixed margin to every video. A claim about an operation requires its progression; one about availability may depend on an earlier warning or later on-screen qualification.

    Google's key moments provide segment navigation through times and labels. They do not certify support for an AI claim. A chapter named after a feature does not establish that the video demonstrates everything the answer attributes to it.

    If the assistant links only to the whole video, record that it supplied no temporal locator. You may add the review interval, clearly identified as the auditor's work. Do not present it as a timestamp generated by the system.

    Compare visuals, audio and text without forcing agreement

    For each claim, record what is visible, what is spoken and what the transcript preserves. The channels do not have to repeat one another: a demonstration can provide visual evidence without narration. Require the channel that the claim actually needs.

    Record componentEvidence to retain
    IdentityVerifiable product, variant and edit
    VisualsInterval, relevant frames and visible feature
    AudioTrack, language, interval and relevant words
    TextOrigin, version, times and exact passage
    ContextConditions, exceptions and necessary continuity
    DecisionAgreements, discrepancies and review limits

    Do not settle disagreements by majority vote across channels. Audio and captions derived from it are not two independent sources. If the text adds a capability neither spoken nor shown, record that difference rather than using it to complete the demonstration.

    The visual product recognition audit helps when identifying the variant is the problem. Seeing a similar product is insufficient to validate a claim about the particular model named in the answer.

    Retain what the video cannot establish too: a visible interface does not by itself establish current commercial terms, and narration does not prove that an operation completed successfully. Distinguish what is shown, what is said and what the reviewer infers.

    Declare sampling and locator limits

    A system may analyze a transcript, selected frames or video and audio. Record the input route and configuration you actually know. Do not assume an assistant with access to a URL played the entire recording.

    Gemini's video documentation describes default static sampling at one frame per second and warns that details in fast scenes may be missed. That is a property of the documented route, not a universal account of assistant behavior. Preserve processing mode, intervals and sampling when exposed.

    A timestamp proposed by a model is a review candidate. Check the passage in the reference edit before accepting it. Do not add milliseconds or frame numbers that the evidence cannot resolve.

    If only excerpts were reviewed, declare which ones and how they were selected. Do not extend the result to the full duration. Failing to find a detail among sampled frames does not establish its absence from the complete video.

    Version corrections without moving the old evidence

    When there is an error, identify what needs changing: caption track, description, demonstration, temporal link or answer text. Do not replace all those materials under one undifferentiated corrected status.

    Preserve before and after versions with dates. If an edit is trimmed or reordered, check affected locators again; retaining the same number of seconds may point to another scene. A new audio track also needs a fresh concordance check.

    The PDF catalog reading audit makes a similar distinction between document and extraction, but its pages and tables cannot replace a video's temporal reference. Keep each evidence type in its own record.

    Do not overwrite a historical review to imply that the corrected edit was always available. If you query the assistant again, retain its answer as a new observation; it does not by itself show that the editorial change caused the result.

    Close with a reproducible record

    Deliver one record per claim-passage relationship: answer, cited URL, edit, track, proposed interval, verified interval, evidence, decision, limitation and owner. Link the necessary files or captures within authorized access and retention rules.

    Before closure, another person should be able to locate the same passage and understand the verdict. If they cannot, check the version, link or alignment. If the evidence remains inaccessible, the status stays unverifiable rather than false.

    This record can accompany your Mentio visibility analysis. Check what each tool retains: the protocol does not assume that Mentio plays videos, transcribes audio or automatically validates temporal citations. The useful result is a reviewable reference, not a guarantee of inclusion in AI answers.

    When the discrepancy concerns the spoken name, the brand names in audio protocol separates transcription, identity and words introduced by recognition hints.

    Frequently asked questions

    Does a YouTube citation prove the AI watched the whole video? No. The visible link does not reveal all internal processing. Record the known route and review the relevant passage in the available edit without attributing an unobservable full playback to the system.

    Is reviewing the transcript enough? Only for claims that the track can support. If a claim depends on visual identity, motion or on-screen text, review those elements and their temporal context too.

    What if the timestamp opens the wrong moment? Retain the original locator and record where it leads. Add the reviewer's verified interval separately, and check whether the edit or the relationship between excerpt and original has changed.

    Can I retain timestamps after editing the video? Do not assume so. Cuts, reordering and speed changes can move the evidence. Identify the new edit and recheck affected intervals without overwriting the previous review.

    Does an inaccessible video count as a false citation? No. Lack of access prevents verification against that evidence. Record the reason and pending work; do not turn an access limitation into a conclusion about unknown content.

    Want to know if AI mentions your brand?

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

    Related articles