Most crypto social-media work optimises for the wrong metric. Follower count is easy to game, easy to inflate, and AI-search systems have stopped weighting it as a signal of authority. They look at engagement rate adjusted for reach, mention sentiment in unbranded category discussion, and consistency of posting cadence. Those are the levers.
Here is the working methodology, with the specific signals we track and the workflow that produces them.
What does AI search actually read as social proof?
When ChatGPT search, Perplexity, Gemini or Google AI Overviews pick sources for crypto-category answers, they weight a few social signals fairly heavily. Not all signals; some are explicitly downweighted (raw follower count, vanity-metrics like “verified” status without engagement to back it up).
The signals that move AI-citation rate in our weekly tracking:
- Engagement rate on recent posts, adjusted for reach and follower base size. Roughly: replies + reposts + bookmarks ÷ followers, normalised to category averages.
- Mention sentiment in unbranded category discussions — what people say when discussing your category without naming you specifically. AI tools cross-reference this against branded discussion to estimate authenticity.
- Posting consistency. Sporadic posting (3 posts a month then silence for 3 weeks) gets weighted lower than steady cadence (12–20 posts a month with regular distribution).
- Founder-led activity vs corporate-account-led. AI tools detect the difference and weight founder-led activity higher in B2B and crypto contexts.
Notice none of those metrics are gameable through paid follower farms or bot replies. They are measurable but resistant to spoofing — which is why AI tools moved to weight them.
Why does outsourced crypto social so often fail?
The pattern we see most often when crypto founders bring us their previous social agency’s output: voice mismatch.
Outsourced posts read like a marketing intern wrote them, because that’s what happened. The followers can tell within 5 posts. Engagement degrades as the audience treats the account as a brand-broadcast feed rather than a founder voice. By month 3 the founder reluctantly takes back the account, posting drops to whatever the founder can manage personally (usually not enough), and the agency invoice continues for the duration of the contract.
The mismatch is visible on the technical level too. AI tools detect founder voice through stylometric signatures: typical sentence length, vocabulary range, opinion calibration, contraction usage, signoff patterns. When the agency’s writer takes over, those signatures shift abruptly within a single account, and the AI extractor downgrades the account’s authority weighting.
How does a real voice intake work?
Two weeks at kickoff, no shortcuts. The deliverable is a voice doc — typically 6–12 pages — that captures vocabulary preferences, sentence rhythm, opinion patterns, technical depth, contraction usage, swearing tolerance, emoji policy, and signoff style.
The intake process: read the founder’s last 200+ public posts across X, LinkedIn, Telegram, and any blog they personally wrote. Listen to the last 10+ podcasts where the founder appeared as guest or host. Pull common words and phrases the founder uses (and the ones they specifically avoid). Capture how they handle disagreement, how they break news, how they engage with critics. The voice doc encodes all of it.
The first 5 posts after the voice doc is signed off go through line-by-line approval with the founder. This is where the doc gets stress-tested. Common adjustments after the first 5: tone calibration on how aggressive to be in disagreement threads; emoji usage (founders’ tolerances vary widely); signoff style; use of self-deprecation. After post 5, voice usually settles and the founder moves to weekly batch approval.
After 90 days, the voice review happens quarterly. Founder voice naturally evolves with company stage; the doc evolves with it.
How is the KOL list actually built?
Crypto KOL discovery is the most-broken sub-discipline in crypto marketing. Inflated follower counts, bot engagement, pay-to-tweet farms that “guarantee” 50 KOLs in a coordinated burst — none of it moves real-buyer attention.
Our scoring is on three signals.
Engagement rate on the last 30 posts, manually verified. Tools that report aggregate ER are easy to game. We sample 30 recent posts, count actual replies (not just likes), check the reply quality (substantive engagement vs single-emoji replies vs spam-like patterns), and compute a clean ER number. KOLs with apparent ER >5% from tools but actual measured ER <1% from this manual sample are excluded. There are a lot of them.
Audience-buyer overlap with the client’s ICP. We sample 100–200 followers per candidate KOL, cross-reference against known buyer cohorts (existing customer LinkedIn profiles for B2B clients, known-active wallet holders for retail crypto clients). If overlap is below 8–12%, the KOL is not in our ICP regardless of follower count.
Content-quality history. Read the KOL’s last 50 posts. Are they actually informed about the category, or do they shill anything that pays? Have they been involved in past pump-and-dumps or rug pulls (yes, we check)? Is the content coherent and on-topic, or does it shift to whatever’s trending? Failed content-quality history excludes the KOL even if the first two scores pass.
The list shrinks dramatically through this funnel. We typically start with 40–60 candidates per niche, end with 6–12. The pricing is transparent and disclosed in the campaign reports.
What does an AMA or Twitter Space that compounds look like?
Most crypto AMAs are one-time events. They produce some engagement on the day, sometimes get clipped into highlights, and then disappear. The ones that compound have a deliberate structural design.
The pattern that works: a 60-minute event built around 8–12 prepared questions in clear Q-format, all of which start with “what”, “how”, “when”, or “why”. The host introduces each question, the founder or guest answers in roughly 2–3 minutes (long enough to be substantive, short enough to be quotable), the host follows up with one clarifying question, then moves on. No rambling. The structure is deliberate because the post-event derivative content (clips, quote cards, summary thread, blog write-up) needs clean answer blocks to extract from.
The AI-citation angle: the same prompts that work in the AMA become content-extractable Q-format blocks when written up. AI tools quote these structures heavily because the format matches their extraction patterns. So a well-structured Twitter Space contributes to AI-citation rate not just on the day, but for the next 6–12 months as the derivative content gets indexed.
We typically run one AMA or Twitter Space per month per active client. The format is recorded, transcribed, edited into a publishable post, and clipped into 4–6 short-form derivatives within 7 days of the event.
What metrics actually predict pipeline?
Three metrics correlate with downstream pipeline in our client engagements; follower count does not.
Branded search volume from social audiences. When social activity is working, it lifts branded search (“[client name]”, “[client name] reviews”, “[client name] vs [competitor]”). We track this in GSC weekly. Sustained 10–20% month-over-month brand-search lift is the marker.
Reply-rate from ICP cohorts. Engagement from your actual buyers, not from random followers, is the leading indicator. We sample replies on key posts and check what proportion came from accounts in the buyer-cohort definition.
Sentiment in unbranded category mentions. When people discuss your category without naming you, what is the sentiment about your space? When the category sentiment trends positive in your competitor-set, it correlates with AI-citation rate uplift on your branded queries inside 30–60 days.
These three are tracked weekly. Vanity metrics — follower count, total impressions, “engagement” without quality adjustment — show up in monthly reports for completeness but do not drive optimisation decisions.
Where does the discovery call usually land?
For crypto founders evaluating SMM, the question is usually whether to outsource at all. The answer is “depends on whether you can spend 2–3 hours on voice intake at the start, plus 30 minutes a week on batch-approval after that”. If yes, outsourcing scales. If no — if the founder cannot or will not engage with the voice doc — the outsourced output will read fake regardless of how good the agency is. We will tell you on the discovery call which side of that line you are on.
The discovery is free, 30 minutes, named lead. Bring a sample of the founder’s writing if you have it; we will pre-read before the call. Worst case, we tell you to stay in-house and save the retainer fee.