×3.8
Allegro sales in 10 months
From 549k zł in January to 2.1M zł in October 2026. That is +1.56M zł of sales a month.
Case study · Analytics and an AI agent network for European marketplaces
AI agents around a business on Allegro and eMAG: sales grew 3.8 times, and ad ROAS rose from 4.1 to 6.5
We built analytics and a network of AI agents for a DIY goods seller. The seller has three Allegro stores in Poland, Czechia, Slovakia and Hungary, stores on eMAG and Kaufland storefronts. Every hour the AI agents collect sales, costs, prices and stock. Our own AI ad agent reviews PPC for every offer each night and prepares a plan for bids and budgets. In 10 months, sales grew from 549k zł to 2.1M zł a month. ACOS fell from 24.6% to 15.5%, and TACOS from 13.6% to 11.2%. A person always makes the decision.
Key figures
+8.1M zł
of sales above the January level in 10 months
13.6M zł in total on Allegro from January to October 2026.
6.5
Allegro Ads ROAS, it was 4.1 in January
ACOS fell from 24.6% to 15.5%. Every 1 zł of ads brings 6.45 zł of ad sales.
11.2%
TACOS: ad share of all sales, it was 13.6% in January
Ad spend grew 3.2 times, and sales grew 3.8 times.
Sales amounts and volumes are changed under NDA. Products, stores and their names are replaced. Percentages, time frames and ratios are real. How we counted
Summary in 30 seconds
With AI agents, the seller’s team
- Sees sales, costs and profit of three stores and all marketplaces in one place, updated every hour
- Sees ROAS, ACOS and TACOS for every store, ad group and offer every day
- Sees the unit economics of every offer: fee, delivery, ads, purchase cost, margin
- Gets a plan for PPC bids, budgets and ad group content from the AI ad agent
- Sees at once where ads got worse, instead of reading long reports
- Knows the month forecast and the gap to the target for every category
- Finds the cause of a sales drop in one review, not in weeks
- Asks the AI assistant questions in plain words and sets tasks by voice in Telegram
What the seller got
In 10 months: sales ×3.8, and ads pay back better
January against October 2026, three Allegro stores. Amounts are changed under NDA, the ratios are real.
- 2.1M zł of sales in October. In January it was 549k zł
- +1.56M zł of sales a month against January
- +8.1M zł of sales above the January level in 10 months
- ROAS 4.1 → 6.5 ad sales per 1 zł of spend, January → July
- ACOS 24.6% → 15.5% spend as a share of ad sales, January → July
- TACOS 13.6% → 11.2% ad spend as a share of all sales, January → October
Want the same growth on your marketplaces? Start with a 60-minute consultation. We go through your ads by ROAS, ACOS and TACOS and show what to hand to AI agents first.
Book a consultationThe challenge
Three stores, four countries, six currencies and no single picture
This is what a week looked like before the AI agents. Each marketplace report lives on its own. You can combine them by hand, but by Monday the numbers are already out of date.
-
Sales
Sales of three stores in three seller panels. Czechia, Slovakia and Hungary in their own currencies. The team counts the total in a spreadsheet.
-
Ads
You see the spend, but not which offer brings money. The ad panel report loses rows: in a large store, a third of the spend disappeared without a split request.
-
Profit
Commission, delivery and ads come in the billing as dozens of codes. Nobody knows which offer makes money after all costs.
-
Prices
The repricer moves prices every hour. Competitors and partners do the same. Only a manual check shows who is cheaper on each listing.
-
Sales drop
Sales fell. There are many reasons: prices, stock, ads, season. Finding the main one by hand takes a long time.
-
Tasks
Team tasks live in chats. Nobody knows what is done and with what result.
With AI agents, the morning looks different. By 03:20 the ad data is ready, and by 08:00 yesterday’s sales and alerts are in the CRM and Telegram.
- Yesterday’s sales
- ACOS by offer
- Competitor prices
- Tasks for the day
People keep the decisions: what to change in ads, prices and the product range.
What the AI agents had to do
- Show sales, costs and profit of all marketplaces in one picture main goal
- Manage PPC ads with real data for each offer main goal
- Find the reasons for sales drops fast very important
- Track competitor and partner prices
- Calculate the month forecast and category targets
- Keep team tasks under control
Solution
A network of AI agents around the stores
An AI agent is a program that collects data on its own, draws conclusions and reports what to do. Some agents run in the cloud every hour. Others run on our Ubuntu server and collect what the marketplace API does not give. All findings come together in one CRM.
Where the AI agents get data
Allegro API
Sales, billing, own prices and offers of three stores, every hour
Ad panel
Spend and revenue for each offer and day. The agent on the server collects it at night
TradeWatch
Prices of all sellers for each EAN, every 2 hours
Accounting system (a 1C equivalent) and supplier
Catalog, stock, purchase prices and orders in transit
eMAG, Kaufland, Amazon
Sales through the API, Kaufland storefronts through the accounting system (a 1C equivalent)
NBP exchange rates
CZK, HUF, EUR, RON, BGN to złoty at each collection
AI agent network
- data agent
- PPC ad agent
- server collection agent
- price agent
- AI assistant
- task agent
What the team gets
Dashboards
Sales, profit, unit economics, categories, ads
Ad plan
Bids, budgets and group composition with the risk to sales
Telegram alerts
Failures, price violations, tasks with a deadline
AI assistant answers
Any question about the data in plain words
Owner and managers
Make the decisions and apply changes in the ad panel
Eight AI agents and what they do
AI data agent
Every hour, it runs 11 stages: sales, costs, own prices, stock, purchase prices, orders in transit, exchange rates. If a store does not respond, it uses the previous snapshot and reports the failure.
AI PPC ad agent
Every night, it gets spend and revenue for each offer. It calculates ACOS, TACOS and ROAS. It finds groups where ads got worse. It prepares a plan for bids, budgets and group composition with an estimate of the risk to sales.
AI collection agent on the server
Runs on our Ubuntu server. Each store has its own browser profile. At night, it collects ads and conversion from the ad panel. Every 2 hours, it builds TradeWatch reports.
AI price agent
Compares the prices of all sellers with the store prices and the brand’s minimum prices. Details are in the brand control case study.
AI analytics agent
Looks for reasons: why sales fell, which offers ran out of stock, which products people buy together.
AI forecast agent
Calculates the month forecast separately for weekdays and weekends, targets for 97 categories and supplier orders in transit.
AI assistant
Answers questions about all CRM data in plain words. It only reads. It hides margin and ads from roles without access. It runs on Claude, a language model by Anthropic.
AI task agent
Takes tasks in Telegram as text and voice. Every day, it reports overdue tasks and does not close a task without a file or a link.
How the AI agents work every day
- STEP 01
At night, they collect ads
From 02:00 to 03:20, the AI collection agent opens the ad panel of each store in turn and gets spend and revenue for 7 days.
- STEP 02
Every hour, they collect sales
The AI data agent gets yesterday’s and today’s data, prices, stock and purchases.
- STEP 03
They calculate
ACOS, TACOS, ROAS, margin after ads, month forecast, gap to target.
- STEP 04
They report
Failures, new price violations and tasks with a deadline come to the CRM notification bell and Telegram.
- STEP 05
They propose
The AI ad agent prepares a plan of changes. The owner checks it and applies it in the ad panel.
AI ad agent
Our in-house AI agent manages PPC ads with data for each offer
Allegro has no open ads API for sellers and no keywords. So ready-made tools see little. Our AI agent takes data directly from the ad panel, calculates each offer and works only with the levers that really exist.
- 1
Collect
Every night: spend, revenue, clicks and impressions for each offer and day. The agent checks the daily total against the ad panel total to the last grosz.
- 2
Calculate
ACOS, TACOS, ROAS and margin after ads. Separately, sales of other products: half of the ad revenue comes from them.
- 3
Find what got worse
The report shows only the groups and offers where ads got worse. Offers without sales used 8–10% of the budget.
- 4
Prepare a plan
Three strategies, a judge and two independent checks. The output: bids, budgets and group composition with a risk to sales of up to 1.5%.
- 5
Apply and check
The owner applies the plan in the ad panel. A week later, the AI agent compares the result with the previous period.
- Before the first review 24 Jun–7 Jul · TACOS 11.8% 849 zł
- After the first review 8–31 Jul · TACOS 9.6% 1,037 zł
- Before the lever plan 7–27 Sep · TACOS 12.4% 809 zł
- After the lever plan 30 Sep–20 Oct · TACOS 10.5% 952 zł
Sales and spend from marketplace billing. The owner applied the plan. Season and prices also had an effect.
Ad metrics
ROAS, ACOS and TACOS by month
The AI agent counts ROAS and ACOS by Allegro Ads panel attribution: the sales that came from ads. It counts TACOS from all store sales. All three are necessary: ROAS and ACOS show the return on ads, TACOS shows their weight in the business.
after the review
4.1 Jan
7.5 Feb
7.2 Mar
7.3 Apr
7.3 May
6.1 Jun
6.5 Jul
At the end of June the budgets went up 2–3 times, and ROAS fell. On 8 July the AI agent reviewed the ads, and in July ROAS rose again.
24.6% Jan
13.3% Feb
13.8% Mar
13.7% Apr
13.8% May
16.4% Jun
15.5% Jul
Allegro Ads statistics exports for three stores, gross cost.
TACOS in October was 11.2% against 13.6% in January. Ad spend grew 3.2 times, and sales grew 3.8 times: ads grow slower than sales.
Growth
Monthly sales: from 549k to 2.1M zł
The AI agent network grew together with the stores. The dashboard went live at the start of the year, the CRM and AI agents in July, our own server in October.
CRM and AI agents
549k Jan
668k Feb
1.0M Mar
1.3M Apr
1.4M May
1.5M Jun
1.6M Jul
1.7M Aug
1.8M Sep
2.1M Oct
Amounts are changed under NDA, proportions are real.
55% 26% 19%
- 55% offers with a price rise on 19 October 493 offers raised to the brand’s minimum price
- 26% offers ran out of stock 99 offers
- 19% other reasons
Periods: 4–17 and 18–31 October 2026.
The AI agent found the main reason in one review. The team knew at once what to fix: partner prices and stock.
Marketplaces
Allegro, eMAG, Kaufland and Amazon in one CRM
Each marketplace has its own API, its own currencies and its own limits. The AI agents bring everything into one picture in złoty.
Allegro every hour
Three stores in Poland, Czechia, Slovakia and Hungary. Sales, billing, own prices, ads and conversion.
- Sales and costs every hour
- Ads and conversion every night
- Competitor prices every 2 hours
eMAG 3 countries
Romania, Bulgaria, Hungary. The eMAG API accepts requests only from allowed addresses, so collection runs from a whitelisted address.
- Sales by day and country
- RON, BGN, HUF to złoty at the NBP exchange rate
Kaufland 7 storefronts
DE, AT, CZ, SK, IT, FR, PL through the accounting system (a 1C equivalent). Orders and actual sale prices.
- Orders by storefront
- Sale prices compared with catalog prices
Amazon SP-API
SP-API access on 11 EU marketplaces through our OAuth bridge. The data is ready for a sales launch.
- Data from 11 EU marketplaces
- Competitor snapshot by category
Accounting system (a 1C equivalent) and supplier every hour
Catalog, stock, purchase prices and supplier orders in transit.
- 1,135 products in the catalog
- Purchase prices without manual entry
TradeWatch every 2 hours
Prices and sales of all sellers for each EAN. The basis for price and brand control.
- 1,085 EANs
- 7,481 offers in each check
Telegram
AI agents write in Telegram when you need to act
Alerts about failures, prices and tasks come to the work chat. You can set a task by voice: the AI agent transcribes the message and creates a card on its own.
6 types of alerts
- Data is more than 3 hours late
- A store did not respond: previous snapshot used
- Group ACOS went up, ROAS below 3
- Task: deadline today
- Task submitted without a result
- New task from a voice message
1 h between data collections
0.00 zł difference between ad data and the ad panel
5 virtual employees with time tracking
AI task agent bot
4 November
📋 Tasks for today
🔴 Overdue (1)
• Update the photos of the “Brush set 5 pcs” listing
It was due yesterday. Assignee: marketplace manager.
⏳ Deadline today (2)
• Reply to the complaint about Seller F’s offer
Can you finish it on time?
• Apply the weekly ad plan
The AI ad agent’s plan is in the CRM.
✅ New task from a voice message
• “Check the stock of 20 l buckets before the weekend”
Card created, due Friday. Assignee: warehouse.
AI assistant
Ask the AI agent in plain words
Not everything fits into dashboards. You can ask the AI assistant a non-standard question. It finds the data on sales, ads, prices and stock on its own and answers in seconds.
- Only reads data. Changes nothing in the stores
- Sees all CRM sections through 25 tools
- Hides purchase prices, margin and ads from roles without access
- An employee can connect their own Claude agent to the CRM with a personal token
AI assistant finds the data on its own
Why did sales fall in the second half of October?
Sales on 18–31 October are 17.4% lower than on 4–17 October.
55% of the drop came from 493 offers with a price rise to the brand’s minimum price on 19 October.
26% of the drop came from 99 offers without stock. The list and the days without stock are on the “Catalog” screen.
Where did ads get worse this week?
In 3 groups. Google “Scrapers and floats”: ACOS 26.7% instead of 18.7%. Search “Brushes”: 22.1% instead of 13.5%.
Both groups have offers without sales for 14 days. The AI ad agent already added them to the weekly plan.
What do people buy most often with a scraper?
A float: it is in 27.9% of baskets with a scraper. This is 3.1 times more often than average.
In 84% of cases, people buy the 350 mm scraper with a float of the same size.
Example with placeholder products. Shares and multipliers come from a real basket analysis.
AI agent findings on screens
All analytics on four CRM screens
The AI agents work inside the CRM that we built for the seller’s process. Products, stores and amounts on the screens are replaced.
01 Dashboard
Sales of three stores and all marketplaces by day and month. Month target and forecast, updated every hour.
02 Unit economics
Each offer: orders, revenue, commission, delivery, ads, purchase cost, profit and margin. You see which product makes money and which one loses money.
03 Ads Analytics
Ads by store, group and offer. By default, you see only what got worse. ROAS is always next to it.
04 Targets and forecast
The month target for each category and the forecast for weekdays and weekends. You see which category is behind and by how much.
Results
Sales grew 3.8 times, ROAS rose from 4.1 to 6.5, ACOS and TACOS went down
We compared January and October 2026, and ads before and after two AI agent reviews.
| Metric | Before | After | Result |
|---|---|---|---|
| Allegro sales a month | 549k zł | 2.1M zł | ×3.8 +1.56M zł a month |
| Sales above the January level | January–October total against 10 months at the January level | +8.1M zł | |
| Ad ROAS | 4.1 | 6.5 | +58% January → July |
| ACOS | 24.6% | 15.5% | −9.1 pp January → July |
| TACOS | 13.6% | 11.2% | −2.4 pp January → October |
| Ad spend a month | 74,800 zł | 235,800 zł | ×3.2 sales grew 3.8 times |
| Pieces sold a month | 83,300 | 314,400 | ×3.8 |
| Offers with sales in a month | 1,044 | 1,865 | +79% |
| Sales per 100 zł of ads, first review | 849 zł | 1,037 zł | +22% daily spend −23% |
| Sales per 100 zł of ads, lever plan | 809 zł | 952 zł | +18% daily sales +28% |
| TACOS after the first review | 11.8% | 9.6% | −2.2 pp |
| Ad data check against the ad panel | Collection on our own server compared with manual export, 3 stores | to the last grosz | |
| Reason for the sales drop | 55% of the drop came from offers with a price rise on 19 October | found in one review | |
Do you want to see your marketplaces the same way? Start with a 60-minute consultation. We review your data and ads and show what to hand to AI agents first.
Book a consultationHow we counted
- ROAS and ACOS: Allegro Ads statistics exports for three stores, gross cost and ad sales by panel attribution. Full months from January to July 2026.
- TACOS by month: ad spend from billing ÷ all sales of the three stores.
- Sales above the January level: January–October sales minus January sales × 10.
- Sales: marketplace payment operations for three stores. We convert sales in Czechia, Slovakia and Hungary to złoty at the NBP exchange rate. January and February are one total per month, from March by day.
- Pieces and offers with sales: data for each offer for each month.
- Sales per 100 zł of ads: sales ÷ ad spend from the marketplace billing × 100. Daily averages over full days.
- Ad comparison windows for the first review (8 July): 24 Jun–7 Jul and 8–31 Jul 2026. For the lever plan (30 September): 7–27 Sep and 30 Sep–20 Oct 2026.
- TACOS: ad spend ÷ all sales. ACOS and ROAS: by the ad panel attribution.
- Sales amounts and volumes are changed under NDA. Products, stores and their names are replaced. Percentages, time frames and ratios are real.
- Other factors: the renovation season in spring and summer, the product range, prices and stock. AI agents are not the only reason for growth. The owner applied the ad plan. So the timing links the plan and the result, but no experiment proves it.
How the project went
From a dashboard to an AI agent network: 10 months
- JANUARY–JUNE
Dashboard
Sales, costs and unit economics of three stores. Collection every night.
- JULY
CRM and AI assistant
CRM in the cloud, collection every 4 hours, purchase prices from the supplier, the first ad review by the AI agent.
- AUGUST
Ads by offer
Ad collection from the ad panel for each offer and day, Ads Analytics, task control in Telegram.
- SEPTEMBER
Virtual employees
Five AI agents with work time tracking, supply forecast, ad plan with real levers.
- OCTOBER
Own server
Ubuntu server, collection every hour, TradeWatch reports every 2 hours, daily conversion.
- NOVEMBER
Agents without manual work
Nightly ad collection for three stores on the server, CRM access for the team’s own agents, AI complaints agent.
What was hard
The ads API is closed
Allegro has no open ads API for sellers. The AI agent gets data from the ad panel in the same session as a person. Every day, it checks the total against the ad panel total.
The 1,200-row wall
The ad panel returns empty pages after ~1,200 rows. In a large store, a third of the spend disappeared without a split request. Now the AI agent splits the request by campaign and group.
Token rotation
Marketplace tokens change at every refresh. Once, a token failure set a store to zero in the report. Now the AI agent uses the previous snapshot and reports the failure.
Marketplace anti-bot protection
The marketplace blocks data center servers. So the AI agents do not scrape pages. They work through the ad panel, the API and TradeWatch.
Ads sell other products
Half of the ad revenue comes from other products. If you pause an offer for low direct ROAS, you lose sales of related products.
What we need from your team
- Access to the marketplace API and seller panels
- Purchase prices or an export from your accounting system
- A person who applies the ad plan and checks the findings
Who it fits
Is an AI agent network right for your business?
It fits if
- You sell on Allegro, eMAG, Kaufland or several European marketplaces
- You have several stores or countries, and you build the overall picture by hand
- PPC ad spend grows, but it is hard to see what it brings
- You have 300+ products
- You want to see the profit of each offer, not only turnover
- Team tasks get lost in chats
You don’t need it if
- You have one store and 20–30 products
- You do not run ads and do not plan to
Questions
What owners usually ask
Which ad metrics does the AI agent count?
ROAS: ad sales per 1 zł of spend. ACOS: spend as a share of ad sales. TACOS: spend as a share of all store sales. The AI agent counts them for every store, ad group and offer every day. By default the report shows only where the metrics got worse.
Does the AI agent change ad bids on its own?
No. Allegro has no open ads API for sellers. The AI agent collects data, calculates and prepares a plan: bids, budgets, group composition. A person applies the changes in the ad panel. The decision stays with you.
How is this better than ready-made analytics tools?
Ready-made tools see only the API. The AI agents also collect data from the ad panel, conversion, TradeWatch, purchase prices and team tasks. And they calculate by the rules of your business.
Do we need to share seller panel passwords?
No. The AI collection agent runs on the project server, in a separate browser profile for each store. The store owner logs in to the panel. The session does not go to third-party services.
How long does the launch take?
In this project, the dashboard went live at the start of the year. The CRM with the AI assistant followed in July, ad collection by offer in August, our own server in October. We start a new project with what hurts most.
Does this work with other marketplaces?
Yes. Besides Allegro, the same CRM has eMAG, Kaufland through the accounting system (a 1C equivalent) and Amazon through SP-API. Each new source becomes one more data agent.
What do employees see?
Each employee has their own role. For example, a viewer does not see purchase prices, margin and ads: the AI assistant removes them on the server.
Can we check the AI agents’ findings?
Yes. The AI agents calculate each figure with clear rules and check it against the marketplace panel. Claude, a language model by Anthropic, understands questions in plain words.
SellerUp · AI automation for e-commerce
A 60-minute consultation on your marketplaces
We review your data, ads and team tasks and show what to hand to AI agents first. You get a written plan within 48 hours. The price is $249.
Where to start
A 60-minute consultation: what to hand to AI agents in your marketplaces
A working session on your stores. We look at your data, ads, prices and team tasks. You get a written plan: what to automate first, in what order and what it costs in hours and money.
- 60 minutes, online
- Estimate in hours and money
- Written plan within 48 hours
Consultation · 60 minutes, online
$249 $399
Credited towards the next piece of work if it starts within 30 days
Book a consultation or leave a request belowConsultation request
What eats the most time right now — yours or your team’s?
- Reply within 1 business day
- We send a meeting time and a payment link
- Honest go / no-go answer
- No “boilerplate proposal”