An AI marketing agency for your e-commerce store: how to choose one

An AI marketing agency for e-commerce is an agency that runs its own layer on top of the ad platforms, connected to data on margin, stock, returns and customer value. It does not optimise for cheap clicks or for reported ROAS, it optimises for profit per order. You can verify one by asking where it sources margin data and how it measures incrementality.

Why this question has changed over the past two years

Not long ago, what you bought from an agency was execution. Someone sat in the account, added keywords, excluded queries and moved bids by hand. Most of that work has now shifted into the platforms' own automation, and it does the job well. Google and Meta both have models that see signals no human in a spreadsheet will ever see.

That changed the question an e-commerce owner needs to ask. Not whether the agency can set up a campaign, but what data it feeds that campaign and how it knows the campaign worked.

A second shift arrived on top of it, in visibility. In February 2026 Ahrefs measured a 58 % drop in click-through rate on queries with AI Overviews, and the share of zero-click searches in the US reached 58.5 % (overview of studies: https://arvow.com/blog/ai-overviews-statistics-2026). Product data, content and the feed have become a shared layer for organic and paid visibility alike. They can no longer be handled as two separate engagements.

What an AI marketing agency for e-commerce does differently from an agency with AI tools

These are two different things, and the SERP blends them into one. The first is AI inside the platform: Smart Bidding, Performance Max, Advantage+. It is excellent at what it does, which is buying ad inventory in real time. It does have one limitation, and that limitation is not a fault. It optimises for exactly what you send it.

The second is an AI layer above the platforms. It does not compete with the bidding, it decides which signal goes into it. It sees the SKU, the purchase price, shipping, payment fees, the return rate and the value of a repeat customer. That same layer then also governs the product feed, titles and creative, because that is where the largest lever sits.

The difference is measurable, not philosophical. Send the platform order revenue and it learns to find orders. Send it profit after returns and it learns to find profit. In most of the stores we have seen, those are not the same products.

What the data says

Optimising for profit is not an agency trick, it is an officially supported feature. Google Ads calculates gross profit as revenue minus COGS and requires the cost_of_goods_sold attribute in the Merchant Center feed (https://support.google.com/google-ads/answer/14943482). Margin reporting only works in combination with Conversions with Cart Data, that is, with cart contents passed at SKU level (https://support.google.com/sa360/answer/13424777). Either it is connected or it is not. You can verify that in ten minutes.

The second layer is returns. The NRF reports that roughly 19.3 % of online orders come back, and 20 to 40 % in apparel (https://nrf.com/research/2025-retail-returns-landscape). Google has conversion adjustments for this, of the RESTATE and RETRACT type, with a 55 day window from the original conversion (https://support.google.com/google-ads/answer/7686447). Without them the algorithm learns on revenue the store never actually had.

The third thing is measuring contribution. The classic eBay experiment published in Econometrica showed that for searches containing the word eBay, roughly 99.5 % of the traffic from paid brand ads would have arrived anyway. In a more recent dataset of 225 geo tests across DTC brands, brand search had a median incremental ROAS of 0.70x, which is below break-even (https://searchengineland.com/incrementality-testing-advertising-winners-losers-442852). Google itself released Meridian, and in May 2026 Meridian GeoX, open geo experiments whose output calibrates the model. That is an admission that attribution alone is not enough to decide a budget.

And a fourth thing, which surprises most people. Analysing roughly 500 campaigns, Nielsen attributed 47 % of advertising's contribution to sales to the creative, against 9 % for targeting (https://www.marketingcharts.com/advertising-trends-230468). So the biggest lever is not the bidding. It is the pace at which you can test creative and product data.

Where we stand

The consensus in the industry says AI needs to be kept on a leash with manual signals, audiences and exclusions, and that the guiding metric is ROAS or CPA. We argue the opposite on both counts, and we have reasons.

First. Constrain the signals less, fix the inputs more. Optmyzr analysed 9,199 accounts and 24,702 Performance Max campaigns. Audience signals are used by 92 % of advertisers and their accounts performed worse across most metrics, search themes are used by 71 % with a largely negative effect, and 58 % of advertisers had the same or better results with no exclusions at all (https://www.optmyzr.com/blog/performance-max-2025-updates-study-analysis/). When a platform gets a clean signal about profit, it does not need someone on the outside guessing the audience for it.

Second. The guiding metric should be profit per order after returns and shipping, not revenue. In the catalogues we have gone through, gross margin between categories routinely differed by a factor of three. Budget then systematically drifts towards products with high revenue and low margin, while the report looks superb. This is the only reason our engine exists as a separate layer above Google, Meta and TikTok. Not because the platforms do anything wrong, but because they know nothing about your purchase price and your returns until we tell them.

Third. Not AI versus humans, but independent measurement above the platform. Meta reports a lift of around 46 % when optimising for incremental conversions, while an independent Haus dataset of 640 incrementality tests over 18 months says the opposite, that 58 % of brands had a higher incremental ROI on manual campaigns and that the automation overstated its own contribution by roughly 12 percentage points (https://news.marketecture.tv/p/meta-advantage-plus-and-incrementality-findings-from-640-tests-2d1d). Both numbers can be true, they simply measure different accounts. Which is why we trust neither without a test on your data.

The numbers we can point to are specific. At the Zlatá Putňa store we saw revenue up 45.8 % and visibility up 574 %, at Jankiv Siblings +583 % engagement, at Domintell +272 % visitors. ROAS up to +300 % is the upper bound we have reached, not an average and not a promise. We got there on a catalogue with good margin data and room to scale budget. On a store where margin is not in the system, we will not promise a result like that.

How to do it: seven steps and five questions

The order of the steps is not arbitrary, each one stands on the one before it.

1. Get COGS into the feed. Google explicitly permits an estimate, for example 80 % of price, just to get the reporting running. You refine the accuracy later.

2. Turn on cart data. Without SKU level data, margin cannot be measured.

3. Connect returns via conversion adjustments. At a 20 % return rate this is the cheapest fix in the whole account.

4. Add customer value. Customer lifecycle goals and New Customer Value mode can attach additional value to a new customer's first purchase (https://support.google.com/google-ads/answer/12080169).

5. Clean up the feed and product data. Titles, attributes, GTIN, images. The same work also lifts organic visibility.

6. Speed up creative. If Nielsen attributes 47 % of the impact to creative, the number of variants tested per month is a real KPI.

7. Measure incrementality. A geo holdout or a user level holdout, at minimum on brand search and on one main channel.

Five questions to put to every agency, ours included. Where do you source my margin data and what exactly is included in it? Can I see into your AI's decisions, meaning is there a log of who changed a budget and why? Who owns the ad accounts, the conversion data and the audiences when the engagement ends? How will you prove the revenue would not have come without the ads? And what does all of it cost, broken out into setup, monthly fee and share of spend?

The fifth question is a test of character. We give you a price on the first consultation, not after three emails. When someone hides it, it is usually not because it is low.

When this does not apply

At low data volume the whole structure is pointless. If a store does a few dozen conversions a month, the model has nothing to learn from and optimising for profit ends up as a more precise measurement of noise. In that case it is better to invest in the product, the feed and the creative, and keep the advertising simple.

If you sell one product with one margin, POAS adds nothing. The entire advantage comes from differences across the catalogue.

Incrementality tests are not free. A geo holdout means switching ads off in part of the market, and for a seasonal store a short test is often meaningless. Recommendations range from a few weeks to three or six months for longer purchase cycles. That is the real price of certainty, and it needs saying up front.

There is also a conflict that no spreadsheet can resolve. Optimising for short term profit pushes towards the narrowest possible targeting, while the Ehrenberg-Bass school shows that growth comes from reaching the whole category of buyers. For a store that wants to double its brand within two years, those two logics collide. We handle it by splitting the budget and admitting that the second part is harder to measure.

And one acknowledged uncertainty to close on. We have no idea how the balance between paid and organic visibility will shift after the next wave of AI results in search. It is equally true that 70 % of PPC managers using AI models report problems with output quality (https://www.searchenginejournal.com/ppc-trends-2026-ai-automation-and-the-fight-for-visibility/558870/). Which is why budget, brand and creative deployment are signed off by a human here. The machine proposes, the human signs.

What to take from this

You do not recognise an AI agency by how many times it writes the word AI. You recognise it by whether it can show you what data its system runs on, which decisions the system makes on its own and which ones a human approves. Everything else is website phrasing.

If you want to know whether the gap between revenue and profit is hiding somewhere in your catalogue, we will go through it with you. A no-obligation 30 minute consultation where we look into the account, tell you what is missing in the data, and estimate the potential. No slide deck and no commitment to continue.

Frequently asked questions

How do I tell whether an agency actually uses AI rather than just talking about it?

Ask about three specific things: what data its system can see (SKU, purchase price, stock, returns, customer value), how often it makes decisions, and whether there is a decision log you can look into. An agency that merely uses AI tools to write copy will fall over on the first question. A verifiable answer takes a minute.

What is POAS and is it better than ROAS?

POAS is return on ad spend calculated from profit rather than revenue. For a store with varying margins it is a better guiding metric than ROAS, because it stops budget drifting towards products with high revenue and low margin. It does not, however, replace a view of the full P&L or the measurement of incrementality.

How much data does a store need for profit optimisation to make sense?

In practice you need a stable volume of conversions for the model to learn from, and a catalogue with differing margins. At a few dozen conversions a month, or with a single product on a single margin, the gain will not cover the cost of the data work. In that case it is better to invest in the feed, the creative and the product.

How can you verify whether advertising really delivered extra sales?

The only reliable proof is an experiment: a geo holdout, a user level holdout, or a marketing mix model calibrated by tests. The numbers in the platform interface tell you who was given credit, not what would have happened without the ads. The eBay experiment showed that roughly 99.5 % of traffic from brand ads would have arrived anyway.

Who should own the ad accounts and the data once the engagement ends?

The client. Ad accounts, conversion data, audiences and learning history belong to the store, not the agency. Moving to a new account means losing the history that automated bidding rests on, and in effect restarting performance. Put account ownership in the contract before the first campaign goes live.

Do I need accurate margin data before we can start?

No. Google itself permits starting with an estimated COGS, for example as a fixed percentage of the product price, so that reporting can get going. The right approach is to refine it gradually for shipping, payment fees, packaging and the return rate. Postponing the project until margins are perfect costs more than starting with an estimate.