We Weight Influence by What We Can Measure, Not by What Moves People

In the largest randomized experiment ever run on human subjects, the indirect effect — influence transmitted person-to-person through friends — was four times larger than the direct effect of the message itself. Almost no measurement system in commercial marketing can see that channel at all. We have built an entire discipline of attribution on the assumption that credit belongs where the click was recorded, and the click is recorded in precisely the places where influence is weakest.

This is the piece I kept trying to fit inside an argument about bots and social listening, and it never fit, because it isn’t really about bots. It’s about a structural bias in how our measurement assigns weight — a bias that synthetic content happens to exploit with unusual efficiency.

The Evidence: Where Influence Actually Lives

Bond et al. (2012, Nature 489:295-298) randomized political mobilization messages across 61 million Facebook users on US Election Day 2010 and validated the outcome against public voting records. The decomposition:

• Direct effect of the message on recipients: approximately 60,000 additional votes
• Indirect effect through social contagion among friends: approximately 280,000 additional votes
• Total: approximately 340,000 additional votes

The indirect effect was roughly four times the direct effect. The transmission ran mainly between close friends. A person who never saw the message became measurably more likely to vote — by about 0.22% per close friend who did see it.

Now ask the practitioner’s question. In a commercial setting, which of those two effects gets credit?

The 60,000. Only the 60,000. The 280,000 has no referrer, no UTM parameter, no impression record, and no identifiable touchpoint. In every last-click model and nearly every multi-touch model in production today, four-fifths of the actual causal effect is invisible.

The Measurability Bias, Stated Formally

This generalizes into something worth writing down, because once you see it you cannot unsee it in a marketing budget.

Let c_i be channel i’s true contribution and o_i ∈ [0,1] be its observability — the share of its true effect that your measurement system can detect and attribute.

Measured contribution: m_i = c_i × o_i

If budget is allocated in proportion to measured contribution, then:

b_i ∝ c_i × o_i

But the allocation that maximizes return is b_i ∝ c_i.

So every channel is misallocated by a factor of o_i relative to the average observability across your mix. Channels with above-average observability are systematically overfunded; channels with below-average observability are systematically underfunded — and this happens even when every individual measurement is perfectly accurate. It is not a data quality problem. It is a selection problem baked into the structure of attribution itself.

The uncomfortable corollary: the better your tracking gets on the channels you can already track, the worse the misallocation becomes, because you are increasing variance in o without touching c.

HOW BIG IS o FOR THE CHANNELS THAT MATTER MOST?

Small. Much smaller than most dashboards imply.

SparkToro ran a controlled experiment with roughly 100 recruited participants generating 1,113 visits across 16 URLs and 11 social networks, checking what analytics actually recorded. Referral misattribution — traffic that arrived from a real source but was logged as “direct”:

• TikTok, Slack, Discord, Mastodon, WhatsApp: 100% misattributed
• Facebook Messenger: 75%
• Instagram DMs: 30%; public LinkedIn posts: 14%; Pinterest: 12%

The authors are explicit that the panel is small and not fully representative, and I’d treat the specific percentages as indicative rather than precise. But the pattern is unambiguous and it is the exact pattern the Bond result predicts: person-to-person channels, which carry the most influence, have observability at or near zero.

You will also see figures claiming 84% or 95% of sharing is “dark social.” Those trace back to a 2016 estimate and I would not repeat them — they’re widely cited and thinly sourced. The SparkToro experiment is smaller but actually measured something.

Where Synthetic Content Fits

Here is the connection to the bot problem, and it’s the reason I think the industry is fighting the wrong battle.

Synthetic content is not primarily competing with your advertising. It is occupying the high-influence, low-observability quadrant — the peer channel — which is both the most persuasive position available and the one your measurement can’t see into.

Automated traffic crossed 51% of all web traffic in 2024, with bad bots at 37% (Imperva Bad Bot Report, 2025). Whoever operates that capacity gets to simulate the channel that Bond et al. showed carries four times the weight of direct messaging, at near-zero marginal cost, in a measurement environment that cannot distinguish it from organic peer influence.

And three well-documented mechanisms convert that occupation into real belief among real people:

Repetition manufactures truth. Fazio, Brashier, Payne and Marsh (2015) found repetition raises perceived truth even for statements that contradict what participants already know. Their title is the finding: knowledge does not protect against illusory truth. The mechanism is processing fluency, misattributed to veracity.

Source discounting decays. Hovland and Weiss (1951) documented the sleeper effect: people discount low-credibility sources at the moment of exposure, then dissociate content from source over time. The skepticism fades faster than the claim.

Volume reads as consensus. Which follows from the first two, and produces genuine opinions in real people that are, by any detector’s definition, authentic.

So the synthetic content problem is not a data hygiene problem sitting in your listening tool. It’s a competitor operating in the channel you don’t measure, exploiting the fact that you don’t measure it.

Trust Is Moving Toward the Channel We Can’t See

This is getting worse rather than better, and the trust data shows the direction clearly.

The 2026 Edelman Trust Barometer (approximately 34,000 respondents, 28 countries, fielded Oct 23 - Nov 18, 2025) found net trust changes:

• National government leaders: −16
• Major news organizations: −11
• Neighbors, family and friends: +11
• Coworkers: +11
• Respondents’ own CEO: +9

Trust is migrating from institutional, addressable, measurable sources toward proximate, interpersonal, unmeasurable ones. Every point of that migration increases the share of true influence sitting in channels with o near zero — which means the measurability bias is widening on its own, without anyone changing their attribution setup.

A Framework: The Influence-Observability Matrix

Sort your channels on two axes — true influence and observability — and four quadrants fall out, each with a different failure mode.

High influence, high observability. Branded search, retargeting, owned email. These are fine, and they are also where almost everyone already over-invests, because measured performance looks spectacular. Much of what they capture is demand generated elsewhere; they harvest, they rarely create.

High influence, low observability. Word of mouth, private sharing, community, employee advocacy, peer recommendation, and — uninvited — synthetic amplification. This is where the Bond 280,000 lives. Chronically underfunded because it cannot defend itself in a dashboard.

Low influence, high observability. Programmatic display in most configurations. Enormous impression counts, meticulous reporting, minimal incrementality. Survives because the reporting is good, not because the channel is.

Low influence, low observability. Genuinely fine to deprioritize. Rare in practice.

Most budget disputes I’ve sat in are actually arguments between quadrant one and quadrant two, conducted with evidence that only exists for quadrant one.

What to Do Instead

1. Measure incrementally, not by attribution

Geo holdouts, matched-market tests, and randomized budget experiments estimate c_i directly and are indifferent to o_i. This is the single highest-value change available, and the reason marketing mix modeling has come back into fashion after a decade of being called old-fashioned. It never stopped being the right instrument; it just wasn’t as flattering.

2. Ask customers, because the channel that can’t be tracked can be reported

Post-purchase “how did you hear about us” surveys are crude, biased, and the only direct read available on the peer channel. Crude and available beats precise and blind. Run them continuously, not as a one-off study.

3. Deliberately correct for o rather than pretending it’s 1

If your model says word of mouth drives 5% of acquisition and you know your tracking cannot see private sharing at all, the honest move is to state a corrected range with your assumption written down. Reporting the uncorrected 5% as a fact is a decision to under-fund it, made silently.

4. Treat unmeasured channels as unhedged exposure

The corollary nobody likes: if a large share of your influence flows through a channel you can’t observe, you also cannot detect an attack on it. That’s the synthetic content risk in a sentence. Triangulate against behavioral series — support tickets, branded search volume, conversion rate — that bots cannot fake.

5. Fund the peer channel structurally, not campaign-by-campaign

Community, advocacy, and referral programs are the only levers that operate inside quadrant two. They will never win a quarterly ROI comparison against retargeting, and they should not be made to compete in one.

Limitations and Honest Caveats

• Bond et al. is political mobilization, not commercial purchase behavior. The 4:1 indirect-to-direct ratio should not be transplanted to your category as a constant. What transfers is the direction and the order of magnitude, not the coefficient.
• The study is from 2010 data on a platform with a very different feed algorithm than today’s. Social transmission dynamics have almost certainly changed; whether toward more or less peer influence is genuinely unclear.
• SparkToro’s dark social experiment is ~100 people. Directionally strong, not a population estimate.
• The observability formalization assumes budget follows measured contribution proportionally. Real allocation includes strategy, politics, and contracts. The bias is a pressure, not a law.
• Incrementality testing has its own serious problems: statistical power, contamination between geos, cost, and the fact that it measures short-run effects on channels whose value is often long-run.
• Edelman’s Trust Barometer is commissioned research from a communications firm. Large sample, published methodology, not peer-reviewed.

Next Steps

• Take your current channel mix and estimate o for each one honestly — not accuracy, observability. Most teams have never separated those two concepts and the exercise is uncomfortable in a useful way.
• Run one geo holdout on your best-performing measured channel. The gap between attributed and incremental contribution is usually the most persuasive internal document you will produce this year.
• Add a single open-ended discovery question to your post-purchase flow. It costs almost nothing and it is the only window you have into quadrant two.

The through-line is simple and it is not really about AI. We measure what is instrumented, we fund what we measure, and influence has never been especially cooperative about being instrumented. What has changed is that something now occupies the unmeasured channel deliberately and at scale. You cannot defend a position you have never mapped — and for most organizations, four-fifths of their actual influence is happening on terrain that has never appeared on the map at all.