The State of Social GTM in 2026:
why volume is dead, and intent is the only edge left.

01 — The collapse of outbound

In 2019, a well-targeted cold email to a director of growth converted to a meeting 6.4% of the time. In 2026, that number is 0.8%. Across the same period, the cost to acquire a B2B SaaS customer has roughly tripled. The math has stopped working.

This isn't a tooling problem. Sales engagement platforms are better than ever. AI personalisation lets a single SDR send ten thousand emails a day, each one referencing the recipient's recent LinkedIn post. And that's the problem. Buyers' inboxes are now a wall of perfectly-personalised noise. They've stopped reading.

So where did the demand go? It didn't. It moved.

02 — The 11.4% finding

For six months we monitored 8.4M conversations across Reddit, X, LinkedIn, Facebook Groups, and Instagram in the B2B SaaS, dev tools, and creator economy verticals. Then we filtered for one specific signal: a person publicly stating they were looking to buy something, before they had contacted any vendor.

The result was startling. 11.4% of B2B buyers post their requirements publicly — usually phrased as "looking for a tool that does X" or "anyone tried Y instead of Z?" — somewhere between two days and two weeks before they fill out a single form. Three years ago, that number was 3.2%.

The buyer's journey moved from inboxes to communities. Outbound is fishing where the fish used to be.

Reddit alone accounts for 42% of these public-research threads in B2B SaaS. X accounts for 24%. LinkedIn, despite the noise, still produces 19% — heavily weighted toward Series A+ buyers. The remaining 15% lives in niche Facebook Groups and Instagram comment threads, which legacy social listening tools largely ignore.

03 — Intent scoring, in plain English

A keyword match is not a buying signal. Someone saying "I love Notion" is not in market. Someone saying "looking for a Notion alternative because the AI features are killing us" almost certainly is. The job of an intent score is to separate the two.

We score every matched thread on a 1–10 scale across four pillars:

Scores ≥ 8.5 are high-conviction; in our dataset they convert to a trial signup 14× the rate of scores under 7. Scores under 6.5 we discard. The model is re-trained weekly per customer using their own outcome data, which means it gets sharper the longer you run it.

04 — Why four parallel agents, not one big one

We tested a single-agent architecture for two months and abandoned it. A generalist model trying to do intent mining, competitor analysis, community discovery, and site auditing simultaneously was worse at all four than four specialised agents running concurrently. Specifically:

• intent-miner — fine-tuned on 1.2M labelled thread-outcome pairs. Optimises for precision over recall; surfacing 50 real leads beats 500 maybes.
• competitor-bot— uses a different model entirely. Crawls AI-search citations (ChatGPT, Perplexity, Gemini) to find who's actually being cited in your space, then maps, strengths and gaps.
• community-scout — ranks the subreddits, X circles, LinkedIn groups, and niche Facebook Groups where your buyers actually gather, so outreach lands where the conversations already are.
• geo-audit — a pure technical crawler. Checks robots.txt, JSON-LD, llms.txt, sitemap freshness, entity presence on Wikidata, and 47 other signals AI search engines use to decide whether to cite you.

The four agents finish at different times — intent-miner in ~4 minutes, community-scout in ~6, competitor-bot in ~8, geo-audit in ~12 — but the user sees results streaming in as soon as the first finishes. Wall-clock from "start mining" to first usable output: under five minutes.

05 — Six-month results, n=312

Between November 2025 and April 2026 we ran the full Social Mine stack across 312 self-selected B2B founders. They ranged from pre-revenue to ~$8M ARR. Median trial-to-paid conversion was tracked against each founder's existing baseline.

The headline: founders running Social Mine on autopilot — meaning replies are sent without human review when intent ≥ 8.5 — saw 47% lower CAC than their pre-Social-Mine baseline at the 6-month mark. The variance was wide (−12% at the 25th percentile to −71%at the 75th), heavily correlated with how sharply the user described their ICP during onboarding. Users who wrote >150 characters of ICP description outperformed users who wrote <50 characters by roughly 3× on every downstream metric.

The single largest predictor of Social Mine outcomes is the quality of the ICP description in step two. Treat onboarding as setup, not chore.

06 — Methodology & honest caveats

Two big honesty notes. One: founders self-selected into this study, which biases for engaged users. We do not claim Social Mine works equally well for everyone — particularly outside B2B SaaS, dev tools, and the creator economy, where our model has the least training data.

Two:the −47% CAC figure compares to each founder's own pre-Social-Mine baseline, not to a control group running cold outbound in parallel. We did not run a randomised trial. Some of the lift may be attributable to founders simply paying more attention to growth during the study window. We think most isn't — the time-to-first-lead and reply-rate figures are not subject to that bias, and they're strong — but we want to be square about it.

Full methodology, the labelled training set, and the per-platform breakdown are available on request for researchers and practitioners.

07 — What's next

Three things we're shipping over the next quarter:

Multi-language intent scoring. Right now the intent-miner is English-only. We have working prototypes for Spanish, French, German, and Portuguese; expect them by Q3 2026.

Vertical-specific models.A B2B SaaS intent-miner and a DTC e-commerce intent-miner score the exact same thread differently. We're splitting the model accordingly.

Outcome-based pricing.If we say we'll cut your CAC by 40% and we don't, you shouldn't pay full price. We're piloting a contingent pricing tier for Series A+ customers in June.

© 2026 Social Mine Research. This paper may be reproduced in full with attribution. Findings, dataset, and replication code are available under CC-BY-4.0 on request. Errors are our own; please email support@socialmine.ai.