06-reference/research

agent native demand generation cac

2026-10-19·research-brief·source: deep-research·by Ray Data Co (deep-research synthesis)
demand-generationagentify-3000cacagent-economicsdistribution

Agent-native demand generation is live at real scale, and nobody has published a controlled customer-acquisition-cost comparison

The question

"What agent-native demand-generation tactics (SEO/content agents, outbound agents, agent-run ad optimization) are actually live in production today, and do they measurably cut customer acquisition cost versus human-run equivalents?"

Raised to adjudicate a live tension in [[01-projects/agentify-3000/opportunities]]: the project scored "find" (demand generation) as the weakest agent-shaped job at 1.9-2.0 of 3, and the founder pushed back on 2026-10-04 that agents can generate demand. That pushback is currently recorded as a caution on distribution rather than a ceiling.

What we already know (from the vault)

What the web says

Convergences and contradictions

Synthesis for RDCO

The primary result is the absence. Across three tactic families, three web searches, three source fetches and a direct attempt at the primary source of the one aggregate dataset everyone cites, I found zero published comparison where the endpoint is an acquired customer, there is a human-run control, and the sample size and time window are both stated. Every CAC percentage in this space traces to a vendor blog citing another vendor blog. That absence is not a gap in my search; it is the state of the evidence, and it answers the question that was asked. Nobody has demonstrated that agent-native demand generation cuts customer acquisition cost, and nobody has demonstrated it does not. What has been demonstrated, repeatedly and with real numbers, is that agents cut the cost per unit of output.

The evidence supports the reframing, and inside the reframing it lands closer to agentify-3000's low score than to the founder's pushback. I am flagging that explicitly per the standing rule on conviction asymmetry, and I am not going to soften it: on production evidence as of today, "find" is weak, and it is weak for the reason opportunities.md already guessed. But the score itself is the wrong instrument. "Find" is not one job. It is produce (write the post, draft the email, generate the creative variant) and earn-attention (land in the inbox, hold the ranking, be trusted enough to get a reply). Agents score near the top of the scale on produce and near the bottom on earn-attention. A mean of 1.9-2.0 is an average over a bimodal split, and averaging a 3 with a 1 to report a 2 hides the only distinction that matters operationally. The defect in the scoring model is the taxonomy, not the number. Splitting find into produce and earn-attention is the change this brief recommends, and it would simultaneously vindicate the founder (agents really can generate demand-generation activity at volume he would be right to call transformative) and preserve the project's finding (the activity does not convert into customers at a lower cost, because the conversion step is gated on trust).

One trend line argues the founder is right on a longer horizon. In the same 100,000-email dataset, AI reply rate moved from 2.8% (2024) to 4.1% (2026) while human moved 4.8% to 5.2%. The gap narrowed from 2.0 points to 1.1 points in two years. If that convergence rate holds, reply-rate parity arrives before 2028, and the "caution, not a ceiling" language in opportunities.md is the correct read on a two-to-three-year horizon even though it is not yet the correct read on today's numbers. That single series is the most load-bearing number in this brief and it deserves re-checking when the study is repeated.

What would have to be true for the low score to flip. Stated falsifiably, any one of these: (1) a published geo-holdout or randomized experiment in which an agent-run acquisition channel delivers lower incremental cost per acquired customer than a matched human-run channel, with at least 1,000 acquisitions, a window of at least eight weeks, and a control run by a party that does not sell the agent; (2) trust-metric parity with a surviving cost gap, concretely agent cold-email inbox placement at or above 85% and AI articles matching human on both top-10 rate and traffic per post, while unit cost stays at least 5x cheaper, since parity on the trust metric plus a surviving cost gap is the exact mechanism by which cost-per-output becomes CAC; (3) a repeat of the 100,000-email study showing AI meetings-booked at or above the human rate (currently 0.7% versus 1.1%). Condition (3) is the cheapest tripwire and the one to watch.

For RDCO specifically, three concrete things follow. First, the current play is correctly specified and the measurement choice is the part the whole industry gets wrong. [[2026-09-29-nov30-credibility-targets]] scores on phData pipeline rather than followers; every CAC claim surveyed here is a traffic, reply or subscriber claim wearing a CAC costume. Keep the pipeline denominator and do not let a cheaper-per-post finding tempt a switch to volume metrics. Second, do not buy agent-run outbound for RDCO at any price. The only matched dataset in existence says agent cold email runs 2.7x the spam-flag rate and 15 points lower inbox placement, and RDCO's entire asset is a named person's credibility attached to a real employer. opportunities.md itself prices deliverability reputation as taking years. A burned sending domain would cost more than the pipeline it could plausibly produce. Third, and this is where the evidence is actually useful rather than cautionary: the thing agents measurably cut is production cost, and the Nov-30 sheet shows 0 of 6 build-in-public pieces shipped. The bottleneck on the credibility play is founder hours converted into published artifacts, which is precisely the cost agents do cut. Agent-native demand generation pays for RDCO as a production subsidy on the credibility play, not as an acquisition channel. That is the same conclusion [[2026-08-24-idle-capital-buying-demand-playbook]] reached about cash, arrived at independently about agents.

One Ray-selected opportunity, awaiting his read. The highest measured-ROI agent category in demand generation is diagnosis, not creation: SEO-audit agents at 11.4x against content-brief agents at 2.9x, and the entire incrementality-testing stack. In agentify-3000's taxonomy that is "judge," not "find," and judge is the job the project did not flag as weak. It also sits directly on the founder's stated domain of agents in production over data infrastructure. An agent-run incrementality or attribution harness, built in public on Snowflake, would be both in-domain and on the one part of demand generation where the evidence says agents win. Flagged as a Ray-selected candidate for the build-in-public queue, not a founder decision he has made.

Why this is in the vault

It adjudicates the open 2026-10-04 tension in [[01-projects/agentify-3000/opportunities]] between the 1.9-2.0 "find" score and the founder's pushback, and it recommends a specific change to the project's scoring taxonomy (split find into produce and earn-attention) rather than a change to the number. It also sets the measurement guardrail for the Nov-30 credibility play in [[2026-09-29-nov30-credibility-targets]]: keep pipeline as the denominator, treat agents as a production subsidy, and do not route RDCO outreach through an agent sender.

Open follow-ups

Related

Sources

Vault

Web (opened or returned in search results this run)