AI-assisted marketing mix modelling estimates how much each channel, including organic search, contributes to sales by analysing aggregated spend, activity and revenue over time. For SEO, it changes the conversation from last-click attribution to incremental contribution, but only if organic search is fed into the model as a proper variable rather than treated as background noise.
Interest in MMM has grown as cookies, app tracking and click-level attribution have weakened. Open-source tools such as Google’s Meridian and Meta’s Robyn, plus commercial platforms, have made it more accessible. The question for organic teams is how their work shows up.
What is marketing mix modelling?
MMM is a statistical model, usually a form of regression, that relates weekly or daily outcomes such as revenue or leads to inputs such as media spend by channel, promotions, prices, seasonality and external factors. It does not need user-level tracking, which is why it survives privacy changes.
The “AI-assisted” part usually means Bayesian methods, automated hyperparameter search, and machine learning used to estimate adstock (how long a channel’s effect lasts) and saturation (where extra spend stops paying back). Some platforms also use language models to explain results in plain text.
Why is SEO hard to model?
Paid channels have a clean input: spend. Organic search does not. There is no weekly budget line that moves visibility up and down, and the effect of content work arrives slowly and unevenly.
- Inputs are indirect: content published, links earned, technical fixes, none of them measured in money
- Effects lag by weeks or months, longer than most paid channel effects
- Organic visibility is affected by things outside your control, such as algorithm updates and competitor moves
- Brand demand drives branded search, so organic clicks partly reflect other channels’ success
If the model simply leaves SEO out, its contribution gets absorbed into the baseline or wrongly credited to whatever paid channel moved at the same time.
How should organic search enter the model?
Most practitioners use an activity or exposure variable rather than spend. Common choices are weekly non-branded organic clicks or impressions from Google Search Console, split from branded queries.
- Use non-branded organic clicks as the main SEO variable
- Model branded search separately, as an outcome influenced by other channels
- Add known algorithm update dates as event variables
- Record major site changes, such as migrations, as events too
- Keep at least two years of weekly data where possible
Treating non-branded clicks as an input is not perfect, because clicks are partly an outcome as well. But it is far better than ignoring the channel, and it gives you a coefficient you can discuss.
What changes for SEO budgets?
Once organic search appears in the model with an estimated contribution, it can be compared with paid channels in the same currency: revenue per unit of activity. SEO investment then competes on evidence instead of on traffic charts.
This cuts both ways. A model may show that a large share of organic clicks are brand-driven and would have happened anyway. It may also show that non-branded organic traffic drives sales at a lower cost than some paid campaigns. Either result is more useful than the usual stalemate between SEO and paid teams.
Does AI search change the picture?
AI Overviews in Google, ChatGPT search and Perplexity answer some queries without a click. In MMM terms, that can reduce organic clicks while brand awareness and later branded searches hold up. If you model only clicks, you may underestimate organic influence.
A practical response is to track AI visibility separately, for example the share of tracked prompts where your brand is cited, and to test it as an additional variable once enough history exists. Treat early results cautiously; the data series are short.
How do you validate the model?
MMM outputs look precise and can be badly wrong. Validation matters more than the tool.
- Check that predicted revenue tracks actual revenue on held-out weeks
- Run geo or time-based experiments for paid channels and compare with model estimates
- For SEO, use content launches or deliberate pauses on a section as natural experiments
- Look at whether coefficients are stable when you add or remove a few months of data
How Alien Road applies AI-assisted marketing mix modelling
In a typical engagement, our team starts with a data inventory: weekly spend by channel, Search Console exports split into branded and non-branded, CRM revenue, promotions and known events. Missing history is the usual blocker, so we document the gaps before any modelling.
We then build a first model with an open-source framework such as Meridian or Robyn, review the results with the client’s paid and organic teams together, and agree which experiments will test the most uncertain estimates. Organic search is always modelled as its own variable, with brand queries separated.
Reporting is quarterly, with the model refreshed and the budget recommendations stated as ranges, not single numbers. Where the model and an experiment disagree, the experiment wins and the model is adjusted.
Which data quality problems break MMM?
Most failed MMM projects fail on data, not on statistics. Spend recorded on invoice dates instead of delivery dates shifts effects by weeks. Revenue that mixes channels, such as marketplace and direct sales, hides what each channel drives. Promotions that are not logged get credited to whatever media ran at the same time.
- Record spend by the week it ran, not the week it was billed
- Separate revenue sources where the business model allows
- Log every promotion, price change and stock-out with dates
- Note tracking changes, such as a GA4 migration, as events
Where to start this week
Export two years of weekly Search Console data and split it into branded and non-branded queries using a regex for your brand names. Put it next to weekly revenue and paid spend in one sheet. Even before any model exists, that table shows whether non-branded organic moves with sales, and it is the first input any MMM project will need.
Expect the first model to be rough. Its main value is often the conversation it forces between paid and organic teams about which data is missing and which assumptions each side has been making.
Keep reading: Advertising · AI Advertising Optimization