Engagement Is a Threshold, Not a Ladder: 10.2M AI Citations Analyzed

GEOforge Research · Citation Study

We pulled every LinkedIn and YouTube citation in the GEOforge database and went and got the engagement numbers. Engagement is a threshold, not a ladder.

By Paris ChildressGEOforge13 min read

On The GEO Show, my co-host Simon Young bet that the social posts AI engines cite are the ones with the most engagement. I bet on relevance to the prompt. So I went into the GEOforge database, pulled every LinkedIn and YouTube citation we have ever recorded, and enriched it with live engagement data. We were both partly right, and the part we were both wrong about is the useful part.

What is the short answer?

Within a single account or channel, a higher-engagement post really is more likely to be cited. Across the whole corpus, the audience size of a cited post is close to irrelevant, and once a post is in the citable pool, engagement tells you nothing about how often it gets cited.

• Simon was right about the within-source tilt. Compared against the same author's other posts, a LinkedIn post with 10× the engagement is 3.6× more likely to be cited.
• I was right about the absolute floor. Nearly half of every YouTube video AI engines cited has under 1,000 views. The median has 16 likes.
• Neither of us saw the ceiling. Citation odds rise sharply out of the bottom two engagement quintiles, then flatten. Being your channel's biggest hit buys nothing extra.

What was measured?

GEOforge has been running prompt sets against ChatGPT, Google AI Overviews and Google AI Mode since April 2026. Every run stores the full citation list. That gives a corpus nobody has to take on faith: 10.2 million citation events across 336,583 unique cited URLs, from 458,276 prompt runs over 848 prompts and 29 brands in eight industries, between 10 April and 16 September 2026.

Social platforms account for 8.8% of all citations in that corpus. That is not a rounding error and it is not a revolution. It is a meaningful minority channel that most B2B brands are currently leaving on the floor.

Which AI engines actually cite social media?

The first thing the data does is settle an argument Simon and I did not know we were having. Social citation is not a property of AI search. It is a property of Google's AI search.

If you are optimising video for AI search, you are optimising for Google. Full stop. And if your buyers live in ChatGPT, YouTube is not your channel no matter how good the video is: LinkedIn out-cites YouTube there by a factor of 74.

Instagram is the cleanest illustration. Google AI Mode cited Instagram 15,905 times. ChatGPT cited it zero times. Not "rarely". Zero. Anyone reporting a single blended share of voice across engines is averaging away the only fact that matters.

How big is the audience on a typical cited video?

I took a stratified random sample of 1,800 cited YouTube videos out of the 9,941 unique videos in the corpus, and pulled live view, like and comment counts for every one. Weighted back to the full population, the typical cited video is small.

LinkedIn tells the same story at a different scale. Of the cited posts I could measure, 60.9% had fewer than 50 total reactions and comments, and 19.5% had fewer than ten. The median cited post has 38. These are not influencer posts. A meaningful share of them are posts nobody noticed.

The second half of this finding is sharper still. Among cited videos, the number of views has no relationship with how often the video gets cited: Spearman's rho is 0.026, with a 95% confidence interval of −0.021 to 0.073. On LinkedIn the correlation between engagement and citation frequency is −0.15 and not significant. Once a piece of content is in the citable pool, engagement stops carrying information about how heavily it gets used.

Does engagement make a post more likely to be cited?

Looking only at cited content is how you fool yourself. The question is not "do cited posts have engagement" but "do cited posts have more engagement than the posts that were not cited". That needs a control group, and the control has to come from the same author, or you are just measuring who has a big audience.

So I built two. On LinkedIn, I sampled 260 authors whose posts appear in our citation data and collected their recent posting history: 4,728 posts from 189 authors, comparing cited against uncited within each author. On YouTube, I took 40 channels from the citation data and pulled their 50 most recent uploads: 2,000 videos, of which 112 had been cited.

Conditional logistic regression with author or channel fixed effects, controlling for content age. The LinkedIn estimate held between 3.30× and 3.86× under leave-one-author-out.

That last row is the one I keep coming back to. On YouTube, once you hold views constant, the like rate does nothing. Whether an audience approved of a video carries no signal. Whether the video was watched at all does.

Which means the mechanism is probably not what either of us assumed on the show. It is not that engines read social proof and reward it. It is that engagement and reach are the same thing on these platforms, and reach is what determines whether a page ends up crawled, indexed, linked and surfaced at all. Engagement is a proxy for existing, not a vote.

Is more engagement always better? No — it stops climbing

If engagement were a ranking signal, more of it would be better all the way up. It is not. Ranking every video against the other uploads on its own channel and splitting into five equal groups, citation odds jump between the second and third quintile and then flatten out.

The honest reading: there is a floor to clear, and above it the curve goes flat. Being your channel's mid-table performer gets you most of the available benefit. Being its breakout hit does not measurably add to it.

The same shape shows up in a simpler test. Take each cited post and ask where it sits in its own author's engagement distribution. If engagement were the gate, cited posts would cluster at the 90th percentile. They do not. Cited LinkedIn posts sit at the 63.6th percentile of their author's own output (95% CI 55.0–74.1) and cited YouTube videos at the 58.3rd (95% CI 50.0–65.2). Both are significantly above the 50th percentile you would see if engagement made no difference. Both are a long way from a gate.

How old is a post when AI engines start citing it?

There is a timing problem with the social-proof theory that I did not expect to find.

A LinkedIn post's engagement life is roughly 48 hours. After that the feed has moved on and the counters stop moving. But citation does not happen in that window. Among the cited posts where we can actually measure it, the median post was 35 days old when we first caught it being cited, and 56.9% were already more than a month old.

So the engine is not reacting to a post that is currently doing numbers. It is retrieving a page whose engagement was settled weeks earlier, and whose reaction count is a static number on an archived page. That is much more consistent with engagement acting as a historical proxy for reach and indexation than with it acting as a live popularity signal.

What should a B2B brand actually do about this?

• Pick the engine before you pick the platform. If your buyers ask ChatGPT, video is not your lever and LinkedIn is. If they hit Google, YouTube outweighs everything else on social by four to one. Reporting a blended share of voice across engines hides the only decision that matters here.
• Clear the floor, then stop optimising for the floor. Get a post past the point where it is seen, indexed and linked. That is a real threshold and it is worth engineering for. Chasing your personal best on reactions is not: the returns flatten well before the top.
• Spend the marginal hour on relevance instead. Engagement moved the within-author odds 3.6×. It explained nothing at all about how often a cited post gets reused. Whatever decides that is on the content side, and my strong prior after looking at this is that it is how precisely the content answers the prompt.
• Publish more, not louder. If a third of cited posts underperform their own author, the winning strategy is more shots on goal at a competent level of engagement, not fewer posts engineered to go big.
• Do not buy engagement. Pods and comment rings inflate exactly the metric that plateaus, and they do it on posts that were never going to answer anyone's question. You would be paying to move from the 60th percentile to the 90th, across a stretch of the curve that is flat.

The obvious next experiment is the one Simon and I committed to on the show: take a prompt with zero share of voice, build a video whose title is that exact prompt and whose script is stuffed with information the internet does not already have, publish it to a channel with no particular audience, and see whether it gets cited. If relevance is doing the work, a 200-view video should be able to win. Based on this data, I think it can. I will report back.

Method and limitations

Citation data. GEOforge production database, 10 April to 16 September 2026. 458,276 prompt runs across 848 prompts and 29 brands against ChatGPT, Google AI Overviews and Google AI Mode, yielding 10.2M citation events over 336,583 unique URLs. Google engines entered the panel on 23 June.

YouTube engagement. Stratified random sample of 1,800 of the 9,941 unique cited video IDs, stratified by citation frequency, with inverse-probability weights applied to all population estimates. Five videos returned no data (deleted or private) and were excluded rather than zeroed.

LinkedIn engagement. 1,451 unique cited posts identified by activity ID. Publish dates decoded from the LinkedIn snowflake ID and validated against 5,639 scraped timestamps (median absolute error 0 seconds). Shares were excluded from the engagement measure because LinkedIn suppresses repost counts on a majority of posts; YouTube comments were excluded for the same reason, since disabled comments are indistinguishable from unused ones.

Control design. Conditional logistic regression with author or channel fixed effects, so every comparison is within a single account. Videos published within 30 days of the panel closing were excluded as structurally ineligible for citation. Channels with no cited video in the observation window drop out of the fixed-effects model, which is why the YouTube model runs on 1,100 videos across 27 channels rather than all 2,000 across 40.

Known limitations. The YouTube control window truncates: each channel's 50 most recent uploads is a narrow slice, and cited videos from those channels that fall outside it have roughly 3.8× the views of the ones inside it. The within-window comparison is internally valid, but the descriptive medians above should be read as properties of the cited population, not of that window. The LinkedIn matched sample skews to lower-volume posters, because an author's cited post has to fall inside their recent history to be measurable; dropping that filter moves the estimate from 3.62× to 3.46×, so the filter deflates the effect slightly rather than inflating it. "Uncited" is an unverified label — a post coded uncited may simply never have been in scope for any prompt we track, which attenuates the measured effects, so the true within-author engagement effects are likely larger than reported. Sample sizes on the matched comparisons are modest (87 cited LinkedIn posts, 112 cited YouTube videos), so the null result on citation frequency rules out strong associations, not weak ones.

Reproducibility. Every figure in this piece was recomputed independently from the raw exports before publication, and three of the original claims were corrected as a result. The corrections are reflected above.

Sources & method. All figures derive from the GEOforge production citation database (10 April – 16 September 2026; 29 brands, 848 prompts, 458,276 prompt runs, 10.2M citation events). YouTube engagement metrics via vidIQ; LinkedIn post engagement via public profile data. Statistical analysis: conditional logistic regression with author/channel fixed effects, stratum-weighted descriptive estimates, and bootstrap confidence intervals. The debate that prompted this study is Episode 33 of The GEO Show with Simon Young.