AI Marketing Hacks 2026.What works. What’s noise.

This is written for a specific person: a founder running £5k–50k a month on paid social, without an in-house creative team, who has noticed that AI tools have changed what's possible but hasn't seen a straight account of what to actually do about it.

It is not a tool list. Tools change every quarter; the mechanics below have not. Where a number appears, it is either sourced at the bottom, our own measured data labelled as ours, or arithmetic you can redo yourself in a spreadsheet.

Start here, because most 2026 marketing advice is built on figures that dissolve when you follow them back. We tried to source each of these and could not.

1. "Creative fatigues after N days"

You'll see ten days, or a peak at days four to seven, or a click-through collapse by day ten. These trace to vendor blogs citing other vendor blogs, with no account count, no spend level, and no stated definition of fatigue. Several contradict each other about when the peak even is.

What's real instead: Meta defines fatigue against your own history, not a calendar. Its Ads Manager marks an ad creative-limited when cost per result exceeds what your past ads delivered, and fatigued when cost per result reaches twice that baseline — counting exposures of the same image or video across your other campaigns too.^[1] That's a definition you can compute on your own account this afternoon.

2. "Keep frequency under 2.5 / refresh at 3.5"

Attributed to Meta constantly. It isn't in Meta's documentation. Meta publishes no numeric frequency threshold; its own guidance says frequency may average one to two per ad set or run much higher depending on budget, audience size and schedule, and treats rising frequency alongside falling performance as a prompt to investigate.^[2] Use frequency as a diagnostic, never as a rule.

3. "The average UK cost per lead is £70"

This one is worth dwelling on as a lesson in how bad the citation chain gets. The figure in circulation is a US dollar cost-per-lead from one vendor's own client base, republished with the dollar sign swapped for a pound sign. The UK edition of that same report contains no cost-per-lead metric at all.

What's real instead: benchmarks are only useful as order-of-magnitude sanity checks, and only when you know whose campaigns they came from. Your own trailing ninety days is a better benchmark than any published average, because it shares your product, your margin and your market.

4. "AI creative lifts performance by X%"

Google publishes real figures for Performance Max asset generation — advertisers using it are 63% more likely to publish with Good or Excellent Ad Strength, Excellent Ad Strength correlates with 6% more conversions, and campaigns with at least one video see 12% more total conversions.^[3] Two caveats nobody attaches: that announcement is from February 2024, not 2026; and Ad Strength measures how completely you filled in your assets, so "generating more assets improves asset-completeness" is close to circular.

The equivalent Meta-side percentages that circulate — 12% higher click-through, 22% higher ROAS, 4% lower cost per result — we could not confirm on any Meta-owned page. The 4% line in particular appears word-for-word across many vendor blogs, which is the signature of one unsourced claim being copied rather than several parties measuring.

5. Compute your replacement rate before you plan anything

Two numbers. C is how many distinct creatives you want live and earning at once. L is their effective life in weeks. The creatives you must ship per week just to stand still is C ÷ L.

Measure your own L rather than borrowing one: take your last 8–12 creatives that carried real spend, note each one's cost per result over its first seven days, find the first week it crosses back above that baseline and stays there, and take the median of the gaps. Use the median, not the mean — one freak winner that ran for five months will otherwise drag your planning into fantasy.

The spread you find is the actual argument for volume. Most accounts have two creatives carrying everything and a long tail that died in a fortnight, and nobody can pick the two in advance. You are buying attempts.

6. Cost the hire properly before comparing it to anything

A 2026 UK recruiter benchmark puts a Head of Paid Media at £75,000–£100,000, averaging £85,000 in London and the South East.^[4] On top of that sit employer National Insurance at 15% above the £5,000 secondary threshold and auto-enrolment pension at 3% of qualifying earnings^[5][6] — about £98,300 all-in, before recruitment fees, tools, bonus, or the three-to-six months between deciding and having output.

Then set that against the replacement rate. One senior hire is one person's throughput; excellent judgement, one pair of hands. The comparison that matters isn't salary against retainer, it's throughput against consumption.

7. Volume is solved; selection is the whole job now

Producing forty concepts used to be the constraint. It isn't. Deciding which four deserve spend — and being right often enough to matter — is now the entire game. Any workflow that generates without a kill step is producing slop at scale, and any vendor selling you volume without explaining the selection is selling you the easy half.

8. Score before you spend, and keep the scores

Put every asset through a consistent rubric before it goes anywhere near an auction, and record the score. The point isn't that the score is truth — it's that a consistent scale makes your losers legible, and legible losers are the only way to learn anything from volume.

Our own data on this is uncomfortable and worth stealing. We scored fifteen of our own launch concepts on the same rubric we use on client work. When we checked which sub-score actually moved the overall: clarity correlated at r=+0.97 and offer at r=+0.91, while hook came in at +0.37, brand fit at +0.05 and distinctiveness at −0.11.

The reason is in the spreads. Hook, brand fit and distinctiveness were near-constant across all fifteen — a house style makes them constant. Clarity ranged 31 points and offer ranged 35. If your creative comes out of one brand system, the hook is your floor, not your lever. What separates your best ad from your worst is almost certainly whether a stranger can tell what you're offering in one read.

9. Separate the maker from the checker

Whatever makes the asset must not be the thing that approves it — that holds whether the maker is a model or a person. The maker always likes its own work. A reviewer who didn't produce it is the cheapest quality mechanism available, and it's the one most AI workflows skip precisely because generation got cheap.

10. Put a flight recorder on the account

Every change to a live ad account should be logged in something a human can read, and reversible. Once tooling is making changes at machine speed, "what changed on Tuesday" stops being answerable from memory. This is unglamorous and it is the single thing most likely to save you a bad month.

11. Test claims, not decoration

Line up your last ten creatives and ask what claim each one makes. If eight make the same claim in different colours, you ran one test ten times. Generation makes decoration nearly free, which means the temptation to vary the cheap thing has never been stronger — and per our own scoring, decoration is the variable that matters least.

1. Compute C ÷ L from your own account. That number decides your creative plan, your budget and who should make your ads.
2. Measure L yourself. Every published fatigue figure we could find was unsourced.
3. Use Meta's real definition — cost per result at 2× your own baseline — not a frequency rule nobody can attribute.
4. Cost a hire at roughly salary + 15% all-in, and remember it buys one person's throughput.
5. Generate at volume, but only behind a scoring gate and a human who didn't make the asset.
6. Optimise clarity and offer before hook. On our own data, hook barely moved the score.
7. Log every account change, reversibly.
8. Test claims, not colours.
9. Distrust any statistic that can't tell you its sample.

If you want the same analysis run against your own account rather than read about, that's what the free audit is: the nine checks, the replacement-rate arithmetic on your real numbers, and a written verdict with the workings shown — including where the evidence wasn't strong enough to be sure.