Measuring AI Content Performance and ROI

Most content teams that adopt AI can tell you how much faster they publish. Very few can tell you whether the programme is worth more than it was eighteen months ago. That gap is the whole problem with ai content roi measurement: velocity is trivially easy to count, and value is not, so the number that gets reported to the board is the one that flatters the tool rather than the one that describes the business.

This page is about closing that gap with the resources a one-to-five-person team actually has. No data engineer. No CDP. A GA4 property that probably has at least one broken conversion event in it, a Search Console account, a spreadsheet, and roughly half a day a month to do the analysis properly.

Start With Cost, Because It’s the Half You Control

Revenue attribution is contested. Cost is not. If you can’t state the fully loaded cost of publishing one asset, before and after AI entered the workflow, you have nothing to divide your results by.

Build the number like this. Take your salary plus employer on-costs (in the UK, NI at 15% above the secondary threshold plus pension plus software plus a share of overhead, so roughly 25-30% on top of gross), divide by realistic productive hours rather than contracted ones. A £45,000 content marketer costs about £56,000 loaded. Divide by 1,600 genuinely productive hours a year, not 1,950, and you get £35 an hour. That’s your unit.

Now cost a single asset honestly:

The interesting line isn’t drafting. It’s editing, which more than doubled. That’s the pattern I see in nearly every team that measures this properly, and it’s why the headline “AI cut our writing time by 70%” translates to a 27% cost reduction in practice rather than a 70% one.

The edit ratio

Track one derived metric religiously: edit ratio = editing hours ÷ drafting hours. Before AI, a competent team sits somewhere around 0.3. With AI drafting, it typically lands between 0.6 and 1.1 depending on how good your prompting and source material are.

Above roughly 0.8, AI drafting has stopped saving money on that asset type. It may still be saving you calendar time, which has its own value, but the cost case is gone. When you see the ratio climb past 0.8 for a specific format (thought leadership and technical explainers are the usual culprits), that’s your signal to pull AI back to research and outlining for that format and leave drafting to a human.

Tag Production Method Before You Measure Anything Else

You cannot compare AI and non-AI performance if you don’t know which is which. Retrofitting this six months later, by memory, produces garbage.

Add one field to your content register, whether that lives in Airtable, Notion, a Google Sheet or custom fields in WordPress. Four values, defined tightly enough that two people would tag the same asset identically:

• 0 — Human: no generative AI beyond spellcheck.
• 1 — AI research and brief: AI used for SERP synthesis, outlining, question mining. Human writes every sentence.
• 2 — AI first draft, heavy human rewrite: AI produced a draft, a human substantially rewrote it, added original data, examples or quotes.
• 3 — AI draft, light edit: AI produced the draft, a human edited for accuracy and tone without adding substantive original material.

Level 2 is where most decent work sits. Level 3 is where the interesting failures cluster, and you want to be able to see that cohort separately rather than having it dissolved into an average.

Alongside it, store: publish date, primary target query, cluster, owning writer, total hours, cost in GBP, and a refresh_date. Seven fields. It takes twenty minutes a month to maintain and it is the spine of everything below. Without a production_method field, your GA4 data is just traffic.

Get the Measurement Layer Wired Properly

There is a persistent fantasy that ai content roi measurement requires a warehouse and a dbt project. For a team publishing under 30 assets a month, it requires GA4 configured correctly, the Search Console API, and a join key.

The join key is the URL path. That’s it. Your content register has a path column, GA4 has pagePath, GSC has page. Everything downstream is a join on that.

Four things worth getting right before you build anything:

Consent Mode v2 is distorting your baseline. Under UK GDPR with a typical cookie banner, expect 20-40% of sessions to be consent-denied. GA4’s behavioural modelling fills some of that back in, but only once your property clears the modelling thresholds (roughly 1,000 daily events with analytics_storage denied for seven consecutive days). Below that threshold, you’re looking at raw consented traffic only. Check Admin → Data Settings → Data Collection, and cross-reference your GA4 organic sessions against GSC clicks for the same period. If GA4 is showing 60% of GSC clicks, you know your multiplier.

Search Console is your truth source for organic discovery. It’s not consent-gated and it’s not sampled the way GA4 Explorations are. Use the GSC connector in Looker Studio for impressions, clicks, position and query data. Use GA4 for what happens after the click.

Turn on the free BigQuery export now, even if you won’t query it for a year. GA4’s UI retains event-level data for 14 months maximum; the export is yours forever. At the volumes a small content team generates, the storage and query cost sits comfortably in single-figure pounds per month.

Define one content-assisted conversion event and stop arguing about attribution models. My preferred definition: a user whose first session landing page was a /blog/ or resource URL, who converted within 90 days on any channel. It’s directionally honest, it’s computable from the GA4 BigQuery export with about 30 lines of SQL, and it doesn’t require you to defend data-driven attribution to a finance director who doesn’t believe in it.

The full build, including the exact GA4 exploration configuration, the Looker Studio blend that joins your content register to GSC and GA4, and the calculated fields for cost-per-click and cost-per-conversion by production method, is set out step by step in Building an AI Content Performance Dashboard in GA4 and Looker Studio. Work through that once and the monthly analysis becomes a 40-minute job.

A Worked Example: B2B SaaS, Two-Person Content Team

Here’s a real shape of programme, anonymised. £4.2m ARR, Manchester, selling to operations managers at mid-market logistics firms. Two content people, one fractional designer, an average sales cycle of 71 days and a £9,400 average first-year contract value.

They introduced AI across briefing and drafting in Q1, ran it for nine months, and measured properly against the preceding nine months.

Read that table slowly, because it’s the most useful thing on this page. Cost per asset fell 30%. Cost per outcome moved by one percent. The saving was consumed almost entirely by lower performance per asset, and the team’s real gain was capacity: they covered 45 more queries and built topical depth in two clusters they’d never have reached otherwise.

That’s a defensible, honest result. It is also nothing like the story the tool vendor tells.

When they segmented by production method, the picture sharpened considerably. Level 1 assets (AI research, human writing) averaged 431 clicks in the first 90 days, above the pre-AI baseline, because better brief quality meant better query coverage. Level 2 averaged 268. Level 3 averaged 94, and eleven of the 29 Level 3 assets never cleared 20 clicks at all.

The response was straightforward: stop producing Level 3 entirely, move that budget into Level 1 briefing for every asset, and accept a lower publishing rate. Nine months after that change, cost per opportunity was £398.

Leading Indicators That Catch Quality Decay Early

Organic performance takes 90-180 days to reveal itself, which is far too slow to be a control loop. You want signals inside three weeks.

Four that actually work:

Engagement rate by production method. GA4’s default engaged-session threshold is 10 seconds, which is too generous for long-form. Create a custom event that fires at 60 seconds and 50% scroll depth, then compare the rate across your production cohorts. A 12-point gap between Level 1 and Level 3 shows up within a fortnight of publishing.

Days to first 100 GSC clicks. Median across a cohort. One team I worked with saw this go from 48 days to 71 days after moving to heavier AI drafting, while their published volume doubled. Volume masked it in the totals; the median exposed it.

Internal link earn rate. How often do writers cite the asset in later pieces? If nobody on your own team finds a post worth linking to, external sites won’t either. Screaming Frog will give you internal inlink counts per URL in a five-minute crawl; export it monthly.

SME rejection rate. If you route drafts past a subject expert before publication, count how many come back requiring substantive correction rather than tweaks. Above 30%, your AI-assisted research layer is producing plausible-sounding wrong things and your editing cost is about to spike.

Refreshes Are Where the Numbers Are Genuinely Good

New-asset ROI from AI is, as above, modest and easy to overstate. Refresh ROI is not, and it is chronically underexploited by teams chasing publishing volume.

The pattern: pull every URL from GSC with impressions above 2,000 and average position between 5 and 15 over the last 90 days. Those are pages Google already trusts that are losing on relevance or freshness. Feed the current page, the top five ranking competitors and the full GSC query list for that URL into your model of choice, and generate a gap analysis rather than a draft. Human rewrites the gaps.

Numbers from one agency’s client programme: 22 pages refreshed, average 4.2 hours each at £35 loaded, plus £300 of tooling, total £3,533. Ninety days later, combined clicks on those URLs were up 11,400 against the pre-refresh period. That’s £0.31 per incremental click, against a Google Ads CPC of £4.80 for the same query set. Even discounting heavily for seasonality and regression to the mean, the gap is not close.

Refreshes also compound with your measurement setup, because you already have a before-and-after on the same URL. No cohort matching required.

Metrics That Will Mislead You

Some numbers are worse than no numbers because they generate confident wrong decisions.

Words published per month measures nothing anyone buys. Time saved, self-reported is reliably inflated by 40-60% against time-tracked reality; if you’re going to claim it, use Toggl or Harvest for a four-week sample rather than asking people. AI detector scores are noise, they produce false positives on human writing at rates above 10%, and Google has never said it penalises AI content as such. Traffic totals hide cohort effects completely, which is exactly what you’re trying to see.

Also be careful with average position in GSC. It’s an impression-weighted average, so publishing a pile of new assets that rank at position 40 will drag your site average down even while your money pages improve. Segment before you panic.

Putting It in Front of Whoever Signs Off the Budget

Your CFO or MD does not want a Looker Studio link. They want three numbers and a decision.

Number one: cost per content-assisted opportunity, this period versus the same period last year. Number two: total content-assisted pipeline value, which is opportunities multiplied by ACV, against total programme cost including salaries and tools. Number three: cost per organic click against the blended paid CPC for the same query set, which is the cleanest like-for-like efficiency comparison available to a content team and usually lands somewhere between 4x and 20x in content’s favour.

Put the production-method breakdown in an appendix. It’s the bit that proves you’re managing the programme rather than just running it, and it’s the bit that justifies the next investment, whether that’s a better research workflow or an additional writer.

Give the analysis a fixed slot. Last working day of the month, 40 minutes, same six views in the dashboard, notes in the same doc. The value of ai content roi measurement comes almost entirely from the consistency of the series, not the sophistication of any single reading, and a slightly crude metric tracked for twelve months beats an elegant one tracked twice.

Start this week with the two cheapest moves: add the production_method field to every asset you’ve published in the last six months, and time-track your next five pieces end to end. Those two datasets will tell you more in a month than any vendor case study, and they’re the inputs everything else on this page depends on.