AI dropped the cost of producing marketing to roughly zero. A blog post, a landing page, fifty ad variants, a week of social — all of it is now minutes of generation, not days of labor. Most people read that as a windfall and respond by producing more. That is the mistake. When production stops being the constraint, producing more is not leverage; it is noise you are adding to a channel already drowning in everyone else's free output. Below a quality bar, more content does not merely fail to help — it actively taxes the trust you have with the people you reach, because you are asking for attention and returning slop.
So the strategic question flipped. It is no longer "how do we make more marketing faster." It is "given that everyone can now make infinite competent-looking marketing for free, what is still scarce?" The answer is taste, trust, and genuine signal — the three things AI does not commoditize. Every tactic below keeps a human at the point of judgment. Every trap removes one. That is the actual line between AI marketing that compounds and AI marketing that rots your reputation.
Be precise about the economics first, because the rest follows from it.
Why "more content" is now negative-sum
When a good was expensive to produce, producing it was a signal in itself — a well-made 2,000-word guide told the reader you had spent real effort, and effort correlated with care and competence. That correlation was the whole reason content marketing worked. AI severed it. The guide now costs nothing to generate, so its existence signals nothing. The reader knows this. Their prior on any given page has shifted from "someone worked on this" to "a machine probably spat this out," and they are right often enough that the prior is rational.
This is why the market for generic content is now negative-sum for the marginal producer. You are competing against an infinite supply of free, equally-competent pages, which drives the value of any single one toward zero, while the volume you add makes the channel worse for everyone including you. The premium relocated — off production, onto the layer that stayed scarce. I've argued the general version at length: when competence is free, taste is the last moat, because taste is the compressed judgment that knows which of a thousand competent options is right, and it is the one input a competitor with your exact tools cannot copy off your screen.
Hold that frame. The DO tactics are the ones where AI does work a human genuinely could not, and a human supplies judgment a machine genuinely cannot. The traps are the ones where you let the machine do the judging.
The tactics that compound
1. Use AI for research and synthesis at a scale a human can't reach
The tactic: Point AI at volumes of source material no human would read, and have it cluster, compare, and surface structure. Feed it 100 competitor landing pages and ask what promises they all make (so you can make a different one). Dump 500 support tickets, sales-call transcripts, and review threads and ask it to cluster the exact language customers use for their problem.
The mechanism: This plays to the one thing the model is unambiguously superhuman at — reading everything and finding pattern across it — while the scarce judgment (which cluster matters, which phrase is the wedge) stays with you. You are not asking it to have taste. You are asking it to be a tireless research assistant so your taste has more to act on.
Concrete example: For Velya, the AI product that qualifies inbound leads for clinics, I had a model read hundreds of real intake conversations and cluster the reasons prospects hesitated. The dominant cluster was not price, as the sales team assumed — it was uncertainty about whether a real person would follow up or they'd be stuck with a bot forever. That single clustered insight rewrote the homepage headline. No human was going to read 400 transcripts and hold the distribution in their head. The model did, and I made the call about what it meant.
The boundary: It fails when you accept the synthesis as the answer instead of the input. The model will confidently cluster noise and hallucinate a theme that isn't there. Treat every output as a hypothesis to verify against raw examples — go read ten actual tickets from the cluster before you believe it. The scale is real; the reliability is not, and the verification is the human job.
2. Draft with AI, then edit until the judgment is yours — never publish raw
The tactic: Use the model to get from blank page to structured draft, then edit so heavily that the load-bearing content — the specific claim, the real number, the example only you have — is entirely yours. The model supplies scaffolding. You supply everything that makes the piece worth reading.
The mechanism: A blank page is a real tax and AI removes it for free. But the model's default output is, by construction, the mean of everything written on the topic — the most average take, phrased in the most average way. Averageness is exactly what a reader who wants your view has no use for. Your edit is where you drag the piece off the mean and onto your actual, defensible position.
Do this / not that:
| Not that | Do this |
|---|---|
| Generate 1,500 words, skim, publish | Generate a 1,500-word draft, keep the outline, rewrite the sentences |
| Let the model's claims stand | Replace every generic claim with a specific one you can defend |
| Ship the model's examples | Cut its invented examples, insert real ones from your own operation |
| Ask "is this good enough" | Ask "does a single sentence here carry information a competitor's blog doesn't" |
The test I use: strike every sentence that could appear verbatim on a competitor's page. If more than a third of the draft survives as generic connective tissue, I'm about to publish the model's mean and I stop. If the AI supplied the load-bearing claims too, that's the real signal — I didn't have anything to say, and editing won't manufacture it.
The boundary: This fails silently when you're writing on a topic where you have no genuine expertise, because then there's nothing to edit toward and you'll rationalize the model's mean as "good enough." The tactic multiplies competence you already have. It cannot create competence you don't.
3. Personalize and segment at a granularity that would be uneconomic by hand
The tactic: Use AI to run personalization and segmentation a human team could never afford at your scale — reworking a message for 40 industry verticals, adapting one core case study into the language of each buyer persona, generating landing-page variants matched to ad intent.
The mechanism: Relevance has always beaten volume, but relevance used to be expensive, so everyone defaulted to one generic message blasted wide. AI collapses the cost of specificity, which lets you clear the relevance floor for many small segments at once instead of one big generic broadcast that clears it for nobody.
Concrete example: Kommerce sells a cash-on-delivery commerce OS, and "trust" means something structurally different to a Casablanca apparel seller than to a Lagos electronics importer — different fraud patterns, different buyer objections, different regulatory texture. Writing 15 genuinely region-specific onboarding sequences by hand was never going to happen. With AI drafting from a spec I wrote for each market and me editing the trust-specific claims, it became a two-day job. Each seller reads copy that sounds like it was written for their market, because in the way that matters it was.
The boundary: Personalization is only a gain while it references something real. Adapting to a buyer's stated industry is relevance. Referencing something they never made public is the creepy trap (see below). The line is whether the input to your personalization is something the recipient chose to make visible.
4. Structure content so machines can extract and cite it
The tactic: Write and format so answer engines — the AI systems that now sit between your content and a growing share of your audience — can lift a clean, self-contained, correct claim from your page and attribute it to you. Direct answers up top, specific claims stated as standalone sentences, real structure (tables, defined terms, explicit numbers).
The mechanism: A rising fraction of buyers never reach your page; they ask an assistant and get a synthesized answer. If your content isn't extractable, you're invisible in that layer no matter how good the prose is. Extractability is a new, learnable craft — the discipline of writing for the extractor so the machine can find, understand, and quote you cleanly. And getting cited by those systems is becoming its own distribution channel, which is the whole argument that GEO is the new SEO: optimize to be the source an answer engine quotes, not just the page a human ranks.
Concrete example: A "how cash-on-delivery reconciliation works" explainer, rewritten so each step was a standalone, factual sentence with the actual failure rates stated inline, started getting surfaced when people asked assistants about COD operations. The generic version — same information, buried in flowing paragraphs — got extracted by nobody, because there was no clean claim to lift.
The boundary: Extractability amplifies reach; it does not create authority. If the extracted claim is wrong, hedged, or generic, being cited hurts you. Structure is necessary and nowhere near sufficient — the claim still has to be true and specific enough to be worth quoting, which sends you right back to tactic 2.
5. Mine what already worked instead of guessing what might
The tactic: Before generating anything new, have AI analyze your own history — which emails got replies, which posts drove signups, which page sections correlate with conversion — and extract the pattern. Then generate from the pattern, not from a blank prompt.
The mechanism: Most AI marketing generates against a generic prior ("write a good subject line"). Generating against your own winners ("here are my 30 highest-reply subject lines, find the structure, write ten more in that vein") replaces the model's mean with your proven signal as the starting point. You're using the model to interpolate within what already works for you, not extrapolate into the average of the internet.
Concrete example: Feeding a model our best-performing Velya outreach and asking it to characterize why it worked surfaced a pattern the team hadn't named: the winners led with a specific observation about the clinic's booking flow, never with a pitch. That became a rule, and the rule improved the next batch more than any amount of blank-slate generation would have.
The boundary: This overfits to the past. If the model only ever generates variations of last quarter's winners, you converge on a local maximum and stop discovering new angles. Use it for the reliable 80%, and deliberately reserve room for human bets the data can't yet justify.
The traps that burn trust
Every trap below shares one structure: it uses AI to remove the human judgment layer and optimize for raw volume or the appearance of effort. That is precisely the move the new economics punish.
Trap 1 — Publishing raw AI content at volume
You generate 200 SEO pages and ship them. They rank nowhere, because search systems now actively demote thin machine content, and they signal nothing, because the reader's prior already assumes a machine wrote them. Why it backfires: you've spent your domain's credibility to add pages to a pile the market values at zero, and the volume makes your site read as a content farm. The tell is that you can't point to a single page and say why it deserves to exist. If you can't, delete it.
Trap 2 — AI-generated "engagement" that produces zero trust
Auto-generated comments, AI-written replies, bot-driven "conversations" — activity that inflates a metric while building no relationship. Why it backfires: engagement was only ever valuable as a proxy for attention and trust. Manufacturing the proxy directly, with AI, produces the number and none of the underlying thing, and audiences now smell automated engagement instantly. You optimized the dial and broke the machine it was measuring. This is the inverse of the truth that word of mouth isn't a channel, it's a consequence — real advocacy is the downstream effect of a product worth talking about, and no volume of synthetic engagement substitutes for the thing that causes it.
Trap 3 — Personalization that crosses into surveillance
"Saw you launched X last week" referencing a public launch reads as attention. "Noticed you've been struggling with Y" inferred from behavioral data reads as being watched. Why it backfires: AI makes deep inference cheap, so it's tempting to personalize on things the recipient never chose to reveal, and the moment they sense that, relevance flips to violation. The line is public and deliberate versus inferred and private. Personalize on what they published, never on what you deduced.
Trap 4 — Faking authenticity everyone can now detect
The mail-merged "I really admired your work on..." note, sent to 10,000 people, each with a machine-filled blank. Why it backfires: the entire value of a personal note was that it was costly — a human spent minutes on you specifically. AI drove that cost to zero, so the note no longer signals what it's imitating, and the pattern is now so familiar that recipients match it to spam in one glance. This is the costly-signal test failing in real time: a signal only works when it's expensive to fake, and cheap fake sincerity is worse than no outreach, because it reads as contempt. If the personalization is cheap for you to produce, it's worthless as a signal — that's not a paradox, it's the definition.
Trap 5 — Optimizing for volume over the relevance floor
Because AI makes output nearly free, the instinct is to maximize it — more posts, more emails, more variants. Why it backfires: every piece below the relevance floor is not neutral; it's a small withdrawal from trust and attention you'll want later. Ten pieces that each barely clear the bar nowhere beat one that's genuinely worth someone's time. The correct optimization target was never volume. It's the floor — the minimum quality below which you don't ship, held constant no matter how cheap production gets.
The through-line: put the judgment where the machine isn't
AI is a competence multiplier, and in a world where competence is free, multiplying it is not an advantage — everyone has the same multiplier. The advantage relocated to the inputs the multiplier can't manufacture: the taste that knows which output is right, the trust that only accrues through costly signals and delivered outcomes, and the genuine signal that comes from a product worth talking about. Those are exactly the layers the traps try to fake and the tactics deliberately preserve.
So the discipline is almost embarrassingly simple. Use AI everywhere it does work a human can't — reading everything, drafting scaffolding, personalizing at scale, structuring for extraction, mining your own history. Keep a human at every point where judgment, trust, or signal is on the line. And hold your quality floor exactly where it was before the cost of production collapsed, because that floor is now the only thing separating you from the infinite free slop you're competing against.
Pick your single highest-traffic AI-generated asset right now and apply the test: strike every sentence a competitor could have published verbatim. If the page doesn't survive, it was never marketing. It was noise you paid to make.