An attribution model for content that compounds over months

The report that said our best content was worthless

We had an article that ranked, pulled steady organic traffic, and got quoted back to us by prospects on sales calls. In the last-click report it was credited with almost no revenue. The deals it influenced converted weeks later through a branded search or a direct visit, and last-click handed all the credit to that final touch.

So by the numbers, our best-performing piece of content looked like a failure. We almost killed the program based on a report that was structurally incapable of seeing what content does. Then we rebuilt how we measure it. Here is the content attribution model we run now and why last-click is the wrong tool for this job.

Why last-click fails content specifically

Last-click attribution gives 100 percent of the credit to the final touch before conversion. For a paid ad that someone clicks and buys from in the same session, that is roughly fair. For content, it is structurally wrong for two reasons.

First, content works on a lag. Someone reads an article, learns something, and does not buy that day. They come back in three weeks through a search for your brand name, or they get added to a sequence and convert two months later. The article did the work; the later touch got the credit. Last-click cannot represent a delay between influence and conversion, so it systematically under-credits anything that plants a seed.

Second, content almost always plays an assisting role rather than a closing one. Its job is to enter the consideration set, build trust, and answer the objection before the call. It rarely sits in the last position because it is not supposed to. A model that only rewards the last position will always rate the assist at zero, no matter how decisive it was.

Put those together and last-click does not just under-measure content. It actively tells you to defund the thing that is working, right at the moment it starts compounding.

The model: lagged windows plus assisted credit

Our model has two halves. The first handles the time delay. The second handles the assist.

Lagged attribution windows

We credit a content touch with influence on any conversion that happens within a lookback window after the touch, including conversions far outside the original session. For organic content the window is 90 days. Someone who read an article in March and converted in May still ties that conversion back to the March read.

The window length is not arbitrary. We measured the actual lag between first content touch and conversion across our own funnel and found a median of 38 days with a long tail past 60. A 7-day or 30-day window would have amputated most of content's real contribution. 90 days captures the bulk of the tail without crediting touches that are too old to plausibly matter.

We stamp every content touch with a timestamp and a person ID at read time, so the lookback is a precise query rather than a guess. When a deal closes, we walk backward 90 days and pull every content touch that person had.

Assisted, weighted credit

Within that window, we do not give content 100 percent like last-click does, and we do not split credit evenly, which over-rewards low-value touches. We use a position-and-decay weighting.

The first touch (how they entered our world) gets a fixed weight because entering the funnel is genuinely valuable. The last touch before conversion gets a fixed weight because it closed. Everything in the middle splits the remaining credit, decayed by recency so a touch closer to the conversion gets slightly more than one further back. Content usually lives in the first and middle positions, which is exactly where last-click gave it nothing.

The weights we run: 30 percent to first touch, 30 percent to last, 40 percent split across the middle with a recency decay. A piece of content that was someone's first-ever touch with us and got referenced again mid-funnel can earn meaningful fractional credit on a deal even if it never appears in the final position.

A concrete example

That article we almost killed. Under last-click it was credited with one small deal in a quarter. We re-ran it through the lagged assisted model over the same period.

The model found that the article was the first touch for 23 people who later became opportunities, and a mid-funnel touch for 31 more. Walking the 90-day windows and applying the weights, it earned fractional credit on 19 closed deals across the quarter. Summed, the fractional credit added up to a meaningful share of pipeline that last-click had assigned entirely to branded search and direct traffic.

The branded search was real, but it was not the cause. People searched our brand because the article taught them who we were. The lagged model could see the seed; last-click could only see the harvest.

How content compounds and why the model has to show it

Content compounds because an article published in January is still pulling first-touch reads in June. Each month it sits in search, it adds new people to the top of the funnel who convert on their own lag. A retainer-style monthly report that resets every 30 days can never show this, because the value is spread across months by design.

Our model rolls forward. The same article keeps earning fractional credit on new deals quarter after quarter as its earlier reads mature into conversions. We can look at a piece a year later and see the full lagged contribution stacked up, which is the only honest way to decide whether content is worth the spend. The decision needs the compounding curve, and last-click flattens it to a single bad number.

The plumbing: how we actually capture the touches

A model is only as good as the data feeding it, and content touches are notoriously hard to capture because most of them happen before a person is identifiable. Someone reads three articles anonymously, then fills out a form on the fourth visit. At that moment we know who they are, but the first three reads were anonymous. If we only count touches after identification, we throw away the top of the funnel, which is exactly where content lives.

We solve this with a stable anonymous ID set at first visit and carried in a first-party cookie. Every content read writes a touch row keyed to that anonymous ID with the URL and timestamp. When the person later identifies themselves, we stitch the anonymous ID to the person ID and inherit the whole prior touch history. The three anonymous reads retroactively become part of that person's path. Without the stitch, the lagged model would start the clock at identification and miss the seed entirely.

We do not rely on tracking parameters in URLs for this, because content gets shared, bookmarked, and found through search, and those parameters fall off. The first-party cookie plus server-side touch logging survives all of that, which is what makes the 90-day lookback trustworthy rather than full of holes.

What the model changed about how we run the program

Honest measurement changes behavior, and that is the real payoff. Three things shifted once we could see content's true contribution.

We stopped killing pieces too early. Under last-click, an article that did not convert in its first 30 days looked dead. Under the lagged model, we wait a full quarter before judging a piece, because its first reads need time to mature into conversions. Several pieces we would have pulled turned out to be among our best once the lag played out.

We started prioritizing first-touch content differently from mid-funnel content. The model tells us which pieces are entry points and which are objection-handlers, because their position distributions differ. Entry-point pieces earn the bulk of their credit in the first-touch position; objection-handlers earn it in the middle. That distinction tells us what to write more of depending on where the funnel is thin.

And we got honest with clients about the timeline. Content is a compounding asset with a slow start, and a model that shows the compounding curve lets us set the expectation up front: this will look flat for a quarter and then accrue. That conversation is far easier to have when the report can actually show the curve six months later instead of resetting to zero every month.

What we would tell another team

If your content reports use last-click, you are reading a metric that is structurally blind to what content does. It cannot see the lag and it cannot see the assist, which are the only two things content provides.

Stamp every content touch with a person and a timestamp. Credit touches across a lookback window long enough to cover your real conversion lag; measure that lag before you pick the number. Weight first and last touches and decay the middle so assists earn fractional credit instead of zero. Then let the model roll forward so a piece keeps accruing credit as its reads mature.

We almost defunded our best content because of a report that could not see it working. The fix was not better content. It was a model honest enough to measure the kind of value content actually creates. More on how we wire attribution at arthea.ai.