Semrush Spotlight: Is YouTube the overlooked channel in your GEO strategy?

An evidence-led YouTube and GEO audit you can actually use, starting with the footage you already have.

If a customer video already contains the answer a potential buyer needs, can anyone find it? Can a search engine understand enough about it to present it usefully? And if an AI system cites it, what does that actually tell us? Kevin Indig’s session on AI search, YouTube and third-party sources at Semrush Spotlight in London on 13 October left me asking those questions about the customer interviews, expert conversations and event recordings I work around.

They pick up where Are you getting the GEO value from all that video? left off. That piece argued that B2B organisations spend heavily for access to a customer or expert, make a polished three-minute film and leave much of the useful testimony in a recording nobody opens again, and that the questions people ask, including in AI search, should shape what evidence is captured before filming. This one is about what happens afterwards: retrieval, evidence and distribution, using footage that already exists rather than commissioning another shoot.

The guide below is how I would approach the first audit. It is a working editorial method. Nothing in it is a tested ranking factor, and nothing guarantees inclusion in ChatGPT or Google’s AI answers.

What Indig’s YouTube findings do and do not tell us

In his Spotlight presentation, Indig drew attention to YouTube’s role as a third-party source in AI-generated answers. My conference notes record a striking emphasis on long-form video and non-branded, category-education questions, and YouTube appearing prominently among third-party citation sources across several client analyses. The sample, definitions and engine mix behind those conference figures were not available to me, so I am treating them as a reason to investigate the channel, not as benchmarks or a licence to tell every B2B team that AI prefers videos of a particular length.

There is also a public basis for taking video seriously. In Winners & Losers of the AI Disruption of Search, a keynote he gave at AirOps Next earlier this year, Indig says YouTube has become the most-cited domain in his data, ahead of Reddit, and that video transcripts had the strongest correlation with being mentioned in an AI answer. The analysis covers more than a million AI answers and over 100,000 citations, but the published summary doesn’t break it down by industry, engine or query type, so I can’t tell how closely it describes B2B finance. Correlation is also the limit of the claim: it doesn’t establish that adding a transcript causes a citation, or that YouTube behaves the same way in every AI system.

The same talk makes a point about trust that matters here. Indig reports that when Google shows an AI Overview, clicks to Reddit and YouTube rise, because people check machine answers against other humans. For a B2B buyer, that human check may well be a practitioner’s or partner’s video rather than the vendor’s own channel.

The more useful questions are narrower. For our audience, in our market, and for the questions that matter to our business: which videos are actually being surfaced? Who owns them? What proof do they contain? Where does our existing footage answer those same questions more precisely?

This is why I would not begin with a video production target or a YouTube subscriber target. I would begin with a small set of questions.

First, separate four things that get mistaken for one result

Traditional video reporting is often dominated by views, watch time, impressions and engagement. Those still matter, but they do not answer the same question as GEO measurement.

For this exercise, I would separate:

Signal	What it tells you	What it does not establish
YouTube discovery	People can find or encounter the video on YouTube	AI systems use it as a source
AI citation	An answer links to the video or its watch page	The brand is recommended or remembered
Brand mention or recommendation	The answer names the business or includes it in a shortlist	The viewer saw the original video or became a customer
Commercial or editorial effect	A qualified enquiry, better buyer understanding, sales use or stronger answer can be documented	That YouTube alone caused the effect

A video may be cited without its publisher being named in the answer. A business may be recommended without a citation to its own site. A useful demonstration may influence someone who never clicks any of the sources.

Indig’s ghost citation research, which compares how four AI engines cite and name brands across 115 prompts in 14 countries, is a reminder not to roll these outcomes into one visibility score. I would record them separately, including the engine, market and date of the observation. One prompt run is a snapshot, not a stable ranking.

The audit: start with ten questions, not a hundred videos

The temptation is to export a full video library and classify every file. That might eventually be useful, but it is a poor first move when the immediate goal is to change a publishing decision.

Take ten real audience questions instead. The quickest places to find them are usually:

sales-call, support and customer-interview transcripts, searched for questions buyers actually ask in their own words;
the search-terms report in YouTube Analytics for videos you already have;
Google Search Console queries for related pages;
YouTube search autocomplete and Google’s “People also ask” boxes;
the follow-up questions AI assistants suggest after a broad category prompt;
the titles competitors and independent creators give their videos.

Then filter. Keep questions where a credible demonstration, practitioner explanation or customer example could genuinely change someone’s understanding, and drop those a short written answer would settle.

For example, a B2B finance-software team might encounter these questions:

What does an accountant actually do before a client’s quarterly reporting deadline?
What does it look like when a small business gets reliable figures sooner?
How does an owner check whether information sent by their accountant has been reviewed?
What changes in a practice when clients supply records late?

These are examples for designing an audit, not claims about which prompts have demonstrable search volume. Their value depends on verification with your audience and data.

For each question, record the intended audience, the useful answer, the sort of proof it needs and the platforms where you expect people to look. The proof category matters as much as the phrase. “Can I see it work?” suggests a product demonstration. “What happened for a business like mine?” may call for a customer account. “What do practitioners actually do?” requires somebody with first-hand experience, not another product montage.

Step 1: Observe the answer environment

Search the same question in the systems relevant to your users: perhaps Google AI Mode or AI Overviews, ChatGPT search and a second assistant. Keep the market and wording consistent as far as the products allow. Do not assume that one response represents everyone else’s experience.

Record:

Is a video cited or recommended at all?
If so, which YouTube channel and which particular video?
Is it a vendor, an independent creator, a practitioner, a publisher or a professional body?
Is the answer about a product, a process, a comparison or a problem?
Does the cited video contain a substantive answer, or does it merely share a title with the question?
Is the brand mentioned, cited as a source, both or neither?

Run each question at least three times per engine, logged out or in a clean profile where possible, on two different days, and save the cited URLs and a screenshot. Indig found that only about 2.2% of citations survived when the same prompt was run three times in a row, and his Consensus Gap analysis reports that 91% of AI citations appear in only one engine. A single run in a single tool tells you very little. I would use this as a diagnostic panel, not an invented ranking table.

Then check the video results directly. Search YouTube for the same question, logged out, and note the top five videos: who published them, how long they are, when, how many views they have, and whether they have chapters, captions and a description that actually answers the question. Do the same in Google’s Videos tab and note any key moments it shows. This is where Indig’s emphasis on long-form video becomes something you can check for your own category rather than assume.

Finally, note whose videos answer the question. If a professional body, an independent practitioner or a partner already owns the useful answer, the best move may be to contribute evidence or an expert to their content rather than to optimise your own channel.

If the results do not feature videos for the selected questions, that is a finding too. It may mean the channel is not the best immediate route for that part of the journey.

Step 2: Match the questions to existing footage

Only now would I open the video inventory. For a first pass, limit it to a dozen or so potentially relevant recordings, whether published on YouTube, embedded on a company site, hosted on another video platform or still sitting in the archive.

Create one row per useful answer, not necessarily one row per finished film. A 40-minute interview might contain two strong explanations and a lot of material that should never leave the archive.

Here is the minimum useful audit record:

Field	What to enter
Buyer question	The specific non-branded question the material genuinely answers
Video and timestamp	Public URL or internal asset ID, with the exact start and end of the answer
Speaker and proof	Who is speaking; observed experience, verified number, demonstration or opinion
Reuse and rights	Public permissions, approvals, confidentiality and whether the material remains accurate
Current destination	YouTube, another host, webpage, article or unpublished archive
Title and description	Does the title describe the answer? Do the first lines of the description say what the viewer will learn?
Captions and chapters	Captions checked by a person (yes/no); chapters present (yes/no)
Watch page	Embedded on a page where the video is the main content; page indexed (yes/no)
Basic facts	Publish date, length, views, channel owner
AI baseline	Engine, prompt, market, date, source/citation, mention and competitors
Decision and owner	Improve, publish, pair with a page, retain, reject; who can actually implement it

For the fastest pass, a spreadsheet is adequate. This does not require a new platform or a permanent evidence-bank project. It needs just enough structure to prevent the same editorial question being asked again next week.

Step 3: Apply a strict evidence test

A good-looking film is not necessarily a good answer. Watch the actual moment and ask: would a person who knew nothing about the brand learn something concrete, specific and trustworthy from this?

A customer saying the product is fantastic is valid testimony about satisfaction, but weak evidence for a detailed process question. A practitioner identifying the step that caused delays and explaining what changed is much stronger for that question. A demonstration showing the process may be more useful still.

I would test five things: relevance to the question, specificity of the answer, credible attribution, current accuracy and permission to reuse it. If one of those is seriously deficient, metadata optimisation will not repair the material.

This is also where editorial judgement has to reject reuse, as I found when routing answers from an event interview: a keyword appearing in a transcript is not evidence that the clip answers the question.

Make the right moment discoverable, not merely the video available

Once you have a genuine answer, there is a practical sequence of improvements. Most are established YouTube or Google Search practice rather than AI-specific tactics.

Give the video a title and description that describe the answer

A title like Customer Story: Acme Ltd is useful to an internal library. It is not very informative to somebody researching an unfamiliar problem. Where the footage genuinely supports it, a title such as How a growing business shortened its monthly reporting process tells the viewer what they will learn.

Keep the promise proportionate to the video. Do not rename a testimonial as a how-to if it never shows the steps, and check whether a customer story’s title was approved by the customer before you change it. In its description guidance, YouTube recommends unique descriptions, useful wording near the beginning and terminology that reflects what viewers are seeking. It also offers search-term information in YouTube Analytics.

A short, useful description should identify the question, who is answering, the evidence or demonstration and the relevant destination for further detail. It should not be an SEO paragraph padded with every conceivable variation of “accounting software”.

Check captions and the underlying words

Captions make video more accessible and help viewers understand what was said. They also give you a way to find the exact moment where the speaker supplies useful evidence. Automatic transcripts are a starting point, not proof of accuracy; check names, numbers, technical terms and regulatory language, then correct them in YouTube Studio or upload a corrected caption file. Captions burned into the picture help viewers but give search engines nothing to read.

Indig’s reported relationship between transcripts and AI mentions is a reason to audit this layer. It is not proof that uploading captions directly increases citations. Treat accurate speech, accessibility, human usefulness and machine-readable context as worthwhile outcomes in their own right.

Where the answer is important, I would also put a concise, attributable explanation in the accompanying written material. That written answer should explain the problem and evidence, not simply announce that the video exists.

Add chapters that name the questions or demonstrations

A long interview can be excellent source material and a frustrating viewing experience if the relevant minute is impossible to find. Chapters make particular sections easier for people to navigate.

For videos hosted on YouTube, add timestamped labels that genuinely reflect what happens at each point. YouTube’s published rules require the first timestamp at 00:00, at least three chapters in ascending order and chapters of at least ten seconds. Google’s documentation explains how YouTube timestamps can also inform video key moments in Search.

A hypothetical chapter sequence might look like this:

00:00 Why reporting was taking too long
02:14 What information was arriving late
05:08 The change to the monthly workflow
08:35 What the owner can now see sooner

Those labels are helpful because they correspond to ideas a viewer may need.

Pair the video with the right page

This is where the earlier EI argument about evidence reuse becomes operational. A video and a written answer can do different jobs without duplicating each other.

The video lets viewers hear a practitioner, inspect a demonstration or see the situation. The page can present the direct answer, clarify the limits, provide accurate numbers, attribute the speaker, link to official guidance where needed and place the example in a wider decision.

For an organisation using a non-YouTube host, the webpage may be the most accessible route to that knowledge. If the goal specifically involves YouTube visibility, uploading to YouTube needs to be an explicit editorial and channel-owner decision, subject to permissions and duplication considerations. A hosted embed and a YouTube watch page are not interchangeable.

For video on your own website, Google’s video SEO guidance covers discoverable embeds, stable URLs, video metadata and Search Console checks. VideoObject and key-moment markup may help Google understand eligible video pages. Those are Google Search features, not evidence that ChatGPT or another assistant will cite the video.

Do not create an extra standalone page for every 40-second clip. Sometimes the correct destination is an existing substantive guide, with a well-placed video and a stronger paragraph.

There is a trade-off. Google’s video guidance only treats a video as eligible for its video features, such as video results and key moments, when it is embedded on a watch page where the video is the main content, and that page is indexed. A clip placed inside a long guide can strengthen the guide without the video itself becoming eligible. If Google video visibility matters for that answer, the YouTube watch page or a dedicated page on your own site does that job, and the guide does the explaining. Either way, check that the pages are open to the crawlers of the AI search tools you care about; OpenAI, for example, says public content must be accessible to OAI-SearchBot to be available in ChatGPT search.

A worked example: one interview, three possible routes

Imagine an existing 25-minute interview with a small-business owner about getting financial information earlier. The finished customer film is already published. It has a branded title, a brief description and captions that have never been checked.

During the audit, three moments stand out.

At 04:20, the owner explains the time previously spent assembling figures and why that delayed decisions. This is potentially useful first-hand evidence for an advice article about getting reliable numbers sooner. The immediate action is to check the specific numbers against the original approved material, then improve the written explanation and the video chapter.

At 11:10, the owner describes working with their adviser when something looks wrong. That may help answer a question about responsibilities, but only if the recording genuinely explains the handoff. If the owner merely says they have a good relationship, it does not earn that route.

At 18:00, the owner praises the product. That may be appropriate customer proof for a product story, but it does not explain how the underlying workflow operates. I would not force it into a how-to page or cut it into another clip simply to increase the asset count.

Now suppose the baseline AI answers cite an independent explainer and a practitioner video, but not the company’s material. The useful hypothesis is not “we need more YouTube”. It is “we may have a specific, under-explained customer answer that belongs in this topic’s evidence ecosystem.”

We can test whether improving its accuracy, metadata and written context changes discoverability, and we should not claim the outcome unless it is observed.

There is also a fourth option the finished film hides. If the full interview is strong, permissions allow it and the speaker agrees, publishing it in full with good chapters may be a better test of Indig’s long-form point than another three-minute edit. The same approach applies to event and practitioner footage.

The first 30 days: one controlled video-GEO experiment

I would not propose a sweeping 90-day channel transformation on the evidence of one conference talk. A first cycle should be small enough to fit inside an existing editorial plan and produce a decision. To make it a fair test, leave one comparable video unchanged as a control, so that any rise in YouTube search impressions can be judged against movement across the channel as a whole.

Stage	What to do	Output
Week 1 — diagnose	Select ten audience questions; sample relevant AI answers and YouTube results; inspect sources and competing videos	Question → source → proof-gap map
Week 2 — audit	Match up to twelve existing videos or answer segments; inspect actual footage, rights, captions and metadata	A shortlist of two or three genuinely reusable answers
Week 3 — improve	With channel and web owners, fix the highest-value titles, descriptions, captions and chapters; pair one answer with the right page	One improved video/page package
Week 4 — assess	Check implementation, YouTube discovery signals and a repeat sample of the same AI questions; document whether anything meaningful changed	Keep, revise or stop decision

A month may be too short for a reliable AI-visibility outcome, so the week-four check is primarily an implementation and early-signal review. Recheck after another four to eight weeks if the question matters enough. There is no defensible promise that an AI engine will refresh its sources on your reporting schedule.

For prioritisation, I would score candidate answers from 0 to 2 for audience need, strength of first-hand proof, current accuracy, rights/readiness and ease of fitting into an existing publication. That is a working triage heuristic, not a validated ranking model. A high score earns inspection, not automatic publication.

The actual success test is whether a real buyer question now has a stronger, accessible answer and whether you can detect any improvement in relevant discovery or use. Keep the measures separate:

YouTube Analytics: impressions and views from YouTube search and from Google, and the search terms that found the video;
Google Search Console: whether the watch page and its video are indexed;
AI answers: whether the video URL is cited and whether the brand is mentioned, by engine, across the repeated sample;
GA4: referrals from AI assistants, treated as a floor, since Indig estimates around 70% of AI traffic is recorded as direct;
sales and discovery calls: a simple question about where the buyer researched the problem;
editorial reuse: whether another team used the evidence.

Most of this needs other people. A rough division of labour:

Task	Usually owned by
Questions, evidence test, title and description wording, chapter labels, transcript corrections, the accompanying page	Editorial
Uploads, metadata changes, caption files, playlists	YouTube channel owner
Watch pages, VideoObject markup, indexing, Search Console	SEO or web team
YouTube Analytics, GA4 AI referrals, AI-visibility tracking	Analytics
Reuse permissions and customer re-approval	Customer marketing or legal

It may turn out that the greatest gain is a better guidance page or a better customer-story brief, with no visible movement in AI citations. That would still be a useful result, provided we do not disguise it as a YouTube ranking win.

When YouTube is the wrong answer

The audit should be allowed to conclude that YouTube is not the route. As a working rule, I would favour YouTube when relevant AI answers and search results in your category already cite or show videos, and the question genuinely benefits from seeing or hearing something. A written page is usually better for factual, numerical or regulatory answers, where people need to check the detail. For evaluative questions such as “which is best for a firm like mine?”, a vendor’s own video is weak evidence, and the more useful work may be with the independent practitioners, publishers and professional bodies whose answers buyers already trust.

The checklist I’d keep next to the planning sheet

Before creating any more video, I would ask:

Which specific audience question is this meant to answer?
Where do relevant AI and YouTube results currently send people for that answer?
Do we have a genuinely useful answer already recorded, and at what timestamp?
Is the evidence attributable, current, accurate and cleared for reuse?
Are the title, first lines of the description, captions and chapters helpful to an unfamiliar viewer?
Does an appropriate written page explain the answer, rather than merely host the embed?
Have we distinguished video discovery, AI citation, brand mention and business impact?
Is there a named channel/content owner and a measurable follow-up question?
If the video cannot answer the question, have we recorded the missing evidence for the next interview rather than creating a superficial derivative clip?

That last point closes the loop with my original argument. The audit is not only about rescuing footage. Sometimes it exposes a question that the recording never answered, and that can make the next interview more useful.

Finding the answer used to be much of the work

There is a practical reason this kind of reuse can remain an aspiration even when the footage exists. Someone has to watch the recordings, identify useful answers, log timestamps, check whether each extract makes sense on its own, write the production brief and connect it with a publishing destination. Experienced editors have always been able to do that. Across an archive, however, it competes with the next interview, article and campaign.

AI can reduce the effort of the first pass by retrieving candidate passages against audience questions and helping prepare briefs and metadata. The opportunity is to make useful editorial work affordable often enough to become a habit. I have not measured the time saving, and extracting a plausible passage is still only the beginning.

In a small practitioner-video experiment, I have now requested three separate edits from an existing interview. The questions concern the work behind a submission, the effect of quarterly deadlines on practice capacity, and clients’ responsibilities for supplying information. These are production requests, not published videos or evidence of improved AI visibility.

The source check also exposed a weakness in the process. The initial proposed cut points ended before some answers had fully concluded. Reviewing the transcript led to extending the extracts so the speaker could finish the thought. A faster selection process is useful only if the final edit preserves meaning; a timestamp suggested by AI needs checking against the recording.

This work could have served an SEO strategy before GEO entered the conversation. What may change is how practical it becomes to recover answers routinely, especially when the system has context about the audience and the pages needing evidence. Production capacity, permissions and publishing ownership still matter: making extraction cheaper does not remove the rest of the work.

That gives this experiment two separate tests. Can AI assistance help turn existing expertise into useful, accurately framed answers with manageable editorial effort? And, once published, do those answers gain relevant discovery or get reused? The first can improve the work before the second produces a measurable result.

The opportunity is not video volume

YouTube matters because it is an important place for people to see demonstrations, hear practitioners and check whether an unfamiliar claim feels credible. Its appearance among sources in AI-generated answers adds another reason to take that material seriously, but it does not remove the need for editorial judgement.

For a B2B team with existing customer, expert and event footage, I would rather optimise three defensible answers to questions that matter than upload thirty clips with artificially question-shaped titles.

That is the next step beyond treating video as content inventory. The material becomes part of a knowledge system: audience questions decide what to investigate; the interview acquires evidence; video and writing make it usable; a focused audit makes it easier to find; and the result improves the next brief.

If an AI system eventually cites it, good. The more durable advantage is that the organisation knows exactly what it can prove, where that proof lives and which question it actually answers.

Reporting and methodology: This guide develops my earlier pieces Are you getting the GEO value from all that video? and The last video should make the next video smarter. The conference observations come from Kevin Indig’s Semrush Spotlight London session on 13 October 2026; see also his AirOps Next keynote and research summary and the Growth Memo analyses linked above, some of which are behind a paywall. The platform guidance is linked to YouTube and Google documentation above. Proposed audit steps and illustrative examples are my editorial recommendations, not tested YouTube or AI-ranking factors. No confidential customer material or private analytics are reproduced here.