Automatic product matching using artificial intelligence

📑 Contents

• What does the platform do and what problems does it solve?
• Why is this platform better than other solutions?
• What you will get after implementation
• Who is it suitable for?
• Quick start in 5 minutes
• Settings: which parameter affects what
• How AI Selects Products: Algorithm and Rules
• Matching window interface
• Operator work scenarios
• Protection from double work and token savings
• Background process notifications
• "Created with AI" mark in links
• Operation log: where to check the status
• Frequently Asked Questions
• Tips for optimal configuration
• Conclusion

What does the Elbuz platform do?

Main functions

• Automatic matching via AI - for each unrelated price list line, finds a suitable candidate in your catalog and associates it with an indication of the confidence level (exact / similar / estimated / not found).
• Two operating modes: auto mode (AI applies everything automatically after importing the price list) and manual verification (the operator looks at the report and confirms the results).
• Smart difference recognition — AI understands that “iPhone 256GB” ≠ “iPhone 512GB”, “black” ≠ “white”, and does not confuse different modifications of the same model.
• Highlighting duplicate keywords — the operator immediately sees what's similar and what's different between the names of two products. It activates on hover and doesn't obstruct the view.
• Manual selection of goods For "not found" rows, the operator can manually select a product from the catalog using smart search or specify a category to add as a new item.
• Mass addition of new products into the catalog with one button - products that are not available are created in a batch linked to the specified category.
• Protection against double work — re-importing the price list does not trigger AI for already verified products, saving up to 90% in tokens.
• Periodic recheck of "not found" — if the catalog is replenished, the AI gets a second attempt after a configurable number of days.
• Filter by keywords — process not the entire price list, but only the required products (for example, “Apple only” or “headphones only”).
• Background process notifications — the operator always knows that the AI is currently processing the price list and will not start manual work in parallel.
• Attribution of links - each connection is marked: "created by AI" or "created by operator", can be filtered and reviewed.
• Export to Excel — one-button connection report with all columns, including the AI flag.

What customer pain points does the Elbuz system solve?

• ❌ "We have 80,000 products and 5 suppliers, and we're drowning in manual reconciliation." - Forget it, AI does it for you.
• ❌ "After import, new items get lost among the old lines; we don't notice them." — the system highlights “not found” separately and offers to add to the catalog with one click.
• ❌ "The operator mixed up the products—associating a black iPhone with a white one." — AI monitors variable features (color, memory capacity, size, brand) and does NOT link different ones.
• ❌ "We don't understand what connections were made by humans and what by AI." - each connection is marked with the flag "created by AI".
• ❌ "We're afraid that AI will make things up." — adjustable confidence threshold (from 70% to 100%); anything below that is checked manually by the operator.
• ❌ "Is it worth re-spending tokens?" — the system remembers which products have already been checked and does not call AI on them when re-importing.
• ❌ "The AI is frozen, and we don't know." — banner and toast notifications show the status in real time.
• ❌ "It's difficult to test tens of thousands of AI solutions." — filters by match type and confidence, statistics, highlighting of matching words.

🏆 Why is this platform better than other solutions?

🎯 What will you get after implementing Elbuz?

1. Up to 95% automation matching routines. The operator confirms only complex cases.
2. Tens of hours of savings per week as an import manager.
3. Clean data in the catalog: every connection is either confirmed by AI or verified by a human, and you know who.
4. Rapid identification of new products — anything that is not in the catalog appears as a separate list with an “Add” button.
5. Exact prices: After contacting, the supplier's price is automatically entered into the product card via the standard Elbuz mechanism.
6. Availability control: Supplier stock - on your store shelf without delays.
7. Supplier Analytics: who is cheaper, who is more stable, who has more overlap with your assortment.
8. Saving on AI LLM tokens: Re-importing the price list does not spend tokens again - the journal is protected.
9. Transparency of operations: everything is visible in the operation log with human-readable messages.
10. Scaling without pain: the system works equally well on 1,000 and 1,000,000 products thanks to the task queue and multithreading.

The main problem you are solving: The gap between "we have 50,000 products in our catalog" and "the supplier has an 80,000-line price list, and we don't know what's ours and what's new." The Elbuz platform closes this gap completely.

👥 Who is the Elbuz platform suitable for?

• Retail stores with a wide range of products (from 5,000 products) that work with several suppliers.
• Marketplaces and aggregators, importing competitors' price lists for price analysis.
• For distributors, where suppliers send articles in different formats.
• For dropshippers, for whom it is important to automatically add new products from suppliers and link them to the storefront.
• B2B platforms, where it is necessary to combine catalogues of different manufacturers.
• For companies with their own production, tracking competitors' prices.
• For any store where “price list → catalog” is not a one-time task, but a daily process.

🚀 Quick start in 5 minutes

1. Import a price list supplier as usual (via standard Elbuz price list import).
2. Open the AI Matching window: in the price list form → “Match” menu (icon with a chain and arrow ▾) → “Match via AI” item.
3. Click "Run AI Matching" The AI will begin processing the items in the background. Progress is visible in the window in real time.
4. Wait until it's finished (Usually a few minutes per 1,000 products—depending on AI LLM load). The "Match Results" table will appear.
5. Check the results Exact matches are already marked automatically, while similar and potential matches are checked by the operator. Click "Apply marked"—the items in the price list are linked to the catalog.

Done. Next, we'll set up automatic mode so it all works without your intervention after each import.

⚙️ Module settings: which parameter affects what

All settings are available in two ways:

• For a specific price: AI Matching window → Settings button (gear).
• Globally for all price lists General system settings → "Price list processing" section → "Product matching." These values are applied automatically if no specific settings are specified for the price list.

1. Automatic mode

What is this: Switch. If enabled, after importing a price list, the AI runs automatically and applies results with confidence above the threshold. If disabled, the operator must click "Run" manually each time.

Recommended: Enable this for all price lists you use regularly. Once set, forget about it.

An important subtlety: The automatic mode does NOT apply "not found" items—they remain for operator review. This is correct: AI shouldn't decide whether to add a new product or not; that's the business's decision.

2. Minimum confidence for auto-application, %

What is this: The threshold ranges from 70 to 100. The AI evaluates each match as a percentage. Anything above this threshold is applied automatically in auto mode. Anything below this threshold is reported to the operator for review.

Recommendations:

• 85% (default) — balance. AI applies precise and confident similar matches, while the operator checks for ambiguous ones.
• 95% — paranoid mode. Only perfect SKU/MPN matches. The operator checks everything else manually.
• 70% (minimum, can't go lower) – maximum automation. Suitable for small and homogeneous catalogs (for example, a single manufacturer).

Attention: It's absolutely impossible to drop below 70%—the AI starts guessing and creates errors. 70 is a hard floor, protecting against bad data.

3. Number of product candidates

What is this: How many products from your catalog the AI retrieves per price list item to select. From 3 to 20.

Recommendations:

• 10 (default) — optimal for most catalogs.
• 20 - if you have a similar product range with similar names (for example, variations of the same model in dozens of colors or memory).
• 3–5 - if the catalog is small and candidates rarely exceed this limit.

Impact on cost: The more candidates, the more tokens per request to the AI LLM. The difference between 10 and 20 is approximately +30% to the cost of a single request.

4. Recheck "not found" in days

What is this: The number of days after which to recheck items for which the AI said "not found." From 0 to 365. The default is 10.

Why is it necessary: Today, the AI decided that the product wasn't in the catalog. A month later, you added this product (or another supplier brought it). Without rechecking, the system will forever skip this line as "already checked." With rechecking, the AI checks again every N days, and if the catalog has been updated, it will link it.

Recommendations:

• 10 (default) — balance. Enough for most stores.
• 3–5 - if the catalog changes quickly (several times a week).
• 30+ — If the catalog is stable, save tokens.
• 0 — Disable. Use this if you want to control rescanning manually (via "Clear" and restart).

🔍 AI algorithm: how exactly does the matching work?

Understanding the steps of the algorithm will help you customize the system to your needs and correctly interpret the results:

1. Selection of goods for processing. The system takes all lines with the following conditions from the price list:
   • are not linked to the base catalog (the price list item does not have a product_id),
   • have a price greater than zero (free lines are usually junk in the price list, we don’t spend money on them),
   • are not in the journal of already processed ones (protection against double work),
   • (optional) are subject to keyword filtering.
2. Search for candidates in the catalog. For each item in the price list, the system calls a smart catalog search (the built-in Elbuz search engine takes into account synonyms, morphology, and article numbers) and finds up to N candidates (the "Number of candidates" setting).
3. Request to AI LLM. If there are candidates, a prompt is generated: "Here is the supplier's product, here are N candidates—choose the most suitable OR answer 'not found'." AI LLM returns the product ID, a confidence score from 0–100, the match type, and a brief justification.
4. Checking the answer. The system verifies that the AI LLM selected an ID from the proposed candidates (protection against "hallucinations"). If the AI LLM "invented" an ID, the result is discarded and marked.
5. Saving the result to the log. If the SKU/MPN match is exact, the line is automatically marked "apply." If it's similar or suspect, it's left for the operator to review.
6. Auto-apply (auto mode only). After processing is complete, the system links all results with a confidence level above the threshold. The link is created using the standard Elbuz mechanism (the same as what the operator does manually in the manual matching window).

Rules that AI follows

The AI is trained to follow strict rules to avoid creating false connections. These rules are hardcoded into the prompt and are not configurable because they are fundamental.

• The manufacturer's SKU is the strongest signal. Exact SKU match = almost guaranteed same item.
• Different model = different products. H540 ≠ H390, X100 ≠ X100s. Even if they are 90% similar.
• Color / memory capacity / size / configuration = different products. iPhone 256GB black ≠ iPhone 512GB white, regardless of the fact that both are iPhones.
• Different brand = always different product. Even if the model is similar.
• If in doubt, "not found". The AI is trained to err on the side of "not found" rather than link incorrectly. This is correct: it's better for an operator to check than for a customer to have the wrong item at the wrong price.

Confidence levels and match types

• 90–100% accurate. SKU/MPN match, as do all variable attributes. Automatically marked for use.
• 70–89% — similar. MPN and brand match, variable attributes match. Auto-applies if the threshold is set within these limits.
• 50–69% — estimated. The name and brand are similar, but the SKU is not confirmed. Manual verification is usually required.
• 0–49% — not found. There's no certainty, it's a separate category. Never automatically applied.

📊 Matching window interface

Header: statistics and control buttons

At the top of the window are cards with statistics on the current price list:

• Total — the total number of results to be checked.
• Accurate — matches by SKU/MPN.
• Similar - high confidence without exact SKU.
• Estimated - average confidence.
• Not found — AI did not find a suitable product.
• Confirmed — marked by the operator (or auto) for use.

Below the statistics are the buttons:

• Run AI matching — Processing begins. The counter in parentheses shows how many items will be processed.
• Apply marked — links all marked lines with catalog products.
• Add new items to the catalog — mass-produces products for "not found" items (visible when such items exist and their category is specified).
• Refresh — reload the results table.
• Clear — complete clearing of the results log for the price list (start from scratch).
• Settings — parameters of this price list (or global).
• Cancel task — appears during processing. Allows you to stop a running AI task.

Filter by keywords

Below the buttons is an input field. Enter a word or several (separated by spaces), and the AI will process only products where that word appears in the name, manufacturer, or category. This is handy when you need to run searches for "headphones only" or "Apple products only."

Results table

Columns:

• ☑ Checkbox — the "apply" mark. Exact matches are marked automatically.
• Supplier's product — name from the price list with article number.
• Candidate from BC — a product from your catalog that the AI considers suitable. For "not found," the text "Not found" and (if specified) the category name.
• Confidence — AI confidence percentage.
• Type — exact / similar / assumed / not found (colored icon).
• Justification — a brief explanation to the AI of why it chose this particular product.
• Actions — icons for the line (see below).

Highlighting matching keywords

Hover your cursor over any line, and identical words in the supplier and candidate product names are highlighted with colored badges. This is convenient for a quick visual check: "Yes, these products are the same" or "No, the color/volume/something else is different." The highlighting appears smoothly, doesn't obstruct your view, and is only visible when you hover over the item.

The backlight algorithm is smart: it preserves the integrity of article numbers (010-02173-03 is not broken into pieces), takes into account single-digit numbers of models (Gen 2, DELTA 2), and ignores noise words (and, for, the, and).

Action icons

• 🔍 Select a product (for "not found" rows) – opens a manual search for the product in the catalog. See scenario 3 below.
• 📁 Specify category (for "not found" lines) - select a category to subsequently add the product to the catalog as a new item.
• ↩ Reset (for rows with a selected product) — unlink, return to "not found." Useful if the AI made a mistake or the operator accidentally clicked the wrong product. Resetting works "in place" — the row is not lost from the list.
• ❌ Delete — remove a line from the results log completely.

👨💼 Operator workflows

Scenario 1: Fully Automatic Operation

Who is it suitable for: You trust AI, import price lists regularly, and want minimal manual labor.

1. In the price settings, enable “Automatic mode” and set the threshold to 85–90%.
2. Import the price list - AI starts automatically in the background.
3. Once a day/week, open the AI-matching window → check “Not found” → add new items.

Scenario 2: Hybrid – AI provides the foundation, operator controls

Who is it suitable for: You want to control the quality of connections, especially for new suppliers.

1. Auto mode is disabled (or enabled, but with a high threshold of 95%).
2. After importing, open the AI Matching window → Run.
3. Wait for processing. Check the statistics in the header.
4. Exact matches are already marked - the operator only checks random ones.
5. Similar and suspected - the operator looks through each one (hover your mouse over it and you'll see matching words highlighted) and ticks them.
6. "Apply marked" - the connections are created.
7. "Add new items" - anything that is not in the catalog is added there.

Scenario 3: Manual selection for complex products

AI marked the product as "not found", but you know that it is in the catalog.

1. In the "Not found" line in the "Actions" column, click 🔍 "Select product".
2. A modal window with a search field will open. The supplier's product name will be automatically entered in the yellow box at the top—you can edit and simplify it, or copy keywords from there.
3. Searches are performed in real time (Umka smart search). Up to 100 results.
4. Click on the desired product - it is linked to the line, and the window closes.
5. If you accidentally selected the wrong one, click ↩ "Reset" in the actions menu. The row will return to "not found" WITHOUT reloading the table (it will not be lost from the overview).

Scenario 4: Adding new products in batches

The supplier's price list contains many products that are not in your catalog.

1. After AI processing, open the report.
2. For "not found" rows, click 📁 "Specify category" and select the catalog category to add the product to (if the category wasn't automatically determined from the price list structure). Once selected, the category name is displayed directly in the row.
3. Once the categories are specified, click “Add new items to the catalog.”
4. All products in a specified category are created in the catalog in a single operation. The link to the price list is also established automatically.

Scenario 5: Re-importing the same price list

You have received an updated price list - the same products + a few new ones.

1. Import the price list.
2. The auto mode starts, BUT it processes only new lines and (according to the schedule) old “not found” ones.
3. Already linked products are not affected—the connection is maintained, and prices are updated as usual through the standard mechanism.
4. Tokens are not re-spent. Budget is under control.

Scenario 6: Manual matching in parallel with AI

The operator opened the manual matching window, and the AI processed the same price list in the background.

1. When opening the window, the operator immediately sees yellow banner: "AI matching in progress. X/Y processed. Manual edits may conflict with AI results. [Open AI window]."
2. A toast notification appears with the same message.
3. The operator can: wait for completion, open the AI window and view the progress, or continue manual work understanding the risk.
4. The banner refreshes every 10 seconds. When the AI completes its work, the banner disappears and a green toast appears, indicating "AI matching complete."

🛡 Protection from double work and token savings

The system leads permanent journal Comparison results. Each price list item is recorded in it only once. When re-launching or importing, the system looks at:

• If the product is already in the journal, skip it (don't spend tokens).
• If the product is not available, we process it.
• If the log shows "not found" and the entry is older than N days (recheck setting), we give the AI a second try.

The link between a product and a price list is identified by a persistent hash of the product's contents (name + SKU)—this hash survives deletion and re-upload of the price list. Therefore, you can safely re-upload price lists—the links and comparison history will be preserved.

When the log is automatically cleared:

• After applying the link (LINK) - the record is deleted because the link is stored in the main mapping table and continues to work.
• After adding a new product to the catalog, the entry is deleted because the product now has an ID in the catalog.
• "Not found" and unconfirmed records remain - they protect against repeated AI calls.

Cost guide: For a price list of 10,000 products, the first run costs approximately the same as processing 10,000 queries to AI LLM. Repeated imports are almost free (we only pay for changed products and outdated "not found" results).

🔔 Background process notifications

The automatic mode runs in the background. To ensure the operator is always aware that the AI is currently processing the price list, a notification mechanism is built into the manual product matching windows:

• Yellow banner At the top of the window: "AI matching is in progress for this price list. N/M (X%) processed. Manual edits may conflict. [Open AI window]».
• Toast notification when opening a window - instant notification.
• Auto-update progress every 10 seconds.
• Toast upon completion — “AI matching complete”, green.

This feature is critical: without it, the operator could start manual matching alongside the AI and cause problems. With the banner, they immediately see whether it's better to wait or open the AI window and check there.

🏷 Mark "created via AI" in the connection log

In the standard Elbuz window “Price list links with the base catalog” (the main window for manual matching), we added a column "Matched by AI" Each connection created by the AI is marked with a flag. This gives:

• The ability to filter only AI connections and review them (for example, if you changed a supplier and want an audit).
• Transparency: you know what the AI did, what the operator did.
• Analytics capabilities: what percentage of connections were created by AI vs. human operators?
• Quality control: If you suspect errors, you can view only the AI connections and check them.

The column is also present in export to XLSX — the report is exported with all the links, and it can be filtered in Excel by this flag.

📋 Operation Log: Where to Check AI Status

All important AI-matching events are logged in the Elbuz "Operation Log." Messages are localized (RU/EN/UA) and human-readable without requiring service codes.

Message types:

• 🔗 Product matching via AI — name of the operation type (visible in the first column of the journal).
• Messages by stages:
  • AI Matching: Task #N created, items to process: 234
  • AI Matching: Queue processing started
  • AI matching completed (job #N): 234 processed, 0 errors. Exact: 156, Similar: 42, Suggested: 12, Not found: 24
  • Auto-applied: 198 linked, 36 requiring review
• "Product Comparison -<название> » — a separate entry for each AI call with the name of the product being processed.
• AI Product Matching - N — summary of the queue for running N parallel tasks.

If something went wrong, the log will help you figure out exactly where. For each operation, you can see the user, time, duration, and color (green for success, red for error).

❓ Frequently Asked Questions (FAQ)

1. What if the AI makes a mistake and links the wrong product?

In the row report, click ↩ "Reset"—the association will be canceled, the row will return to "not found," and you can select the product manually. If the error has already been applied and is in the association log, open the main manual matching window, find the association (you can filter by the "AI-generated" flag), and delete it. The next time the AI runs, it will try again (or simply fail to associate if you're unsure).

2. Is it possible to run comparisons on specific products rather than the entire price list?

Yes – enter keywords in the filter. The system will only process products where the keywords are found in the name/manufacturer/category/price list category. OR logic – any word in any field.

3. How much does matching cost in AI LLM tokens?

On average, one product = one request to AI LLM. The request size depends on the number of candidates (the "Number of candidates" setting). With 10 candidates and the gpt-5.4-mini model, the cost is approximately $0.001 per product. For a price list with 10,000 rows, the cost is approximately $10. Re-importing the same price list is almost free (the log protects against duplicate work).

4. What if the supplier has more products than in my catalog?

This is normal. The AI will find matches for those that exist and mark the rest as "not found." The operator then decides whether to add them to the catalog or ignore them. You can bulk-create them in the catalog using "Add New Items."

5. If I turn off auto mode, will my data be lost?

No. The results log and connections are saved. It will simply stop launching automatically—you'll need to open the window and click "Run."

6. Can AI be used for competitors' (not suppliers') prices?

Yes. Linking a competitor's product to your catalog is the foundation of price analysis. AI does this equally well for any price list type. After matching, you get a ready-made "your product → competitor's price" table and can build reports, pricing strategies, and alerts.

7. What should I do if the AI returns a lot of "not found" results, but I know that the products are available?

Check:

• Is the product catalog filled in? Manufacturer's SKU (MPN). This is the main signal for AI.
• Is it indicated? brand for products in the catalogue and price list.
• Are there too few candidates (try increasing the setting to 15-20).
• Isn't the name in the catalogue too "poor" (for example, "Article 12345" instead of the real name).

A good catalog = good AI results. This rule is universal.

8. Is it possible to “cancel” an already running automatic matching?

Yes. In the AI Matching window, when a task is running, a "Cancel Task" button appears. When canceling, you can keep the already processed results (recommended) or delete them (if you want to start over).

9. Does the global mode—for all price lists at once—work?

Currently, no. AI matching works on a specific price list. If you open a window without a price list selected, you'll see the message "Open from a specific price list." Auto mode serves the same purpose in practice: immediately after importing a price list, AI is launched independently on it.

10. Can AI process products without a price?

No, such lines are automatically skipped. Products without a price in the price list are usually junk (empty lines, headings, separators), and there's no point in spending tokens on them.

11. How often does AI recheck "not found"?

According to the "Recheck 'not found' entries after (days)" setting, the default is 10. Each run (manual or automatic) checks: if there is a "not found" entry in the log older than N days, it is rechecked (in case the catalog has been updated). 0 = rechecking is disabled.

12. What if I want full control over each connection—no auto-apply at all?

Disable auto mode in the settings. Then the AI will only launch when the "Run" button is pressed. The results will be stored in the report; nothing will be applied automatically. The operator must click "Apply Marked."

13. Are AI results saved between sessions?

Yes. The results log is permanent (stored in the database). You can close the window, come back tomorrow, and reopen it—all the data is there. Task progress is also saved: if the operator opens the window while the AI is still running in the background, the operator will connect to the existing process.

💡 Tips for optimal configuration

1. A well-filled catalogue is half the battle. Make sure your products have the manufacturer's stock number (SKU), brand, and a clear name. AI works with what's available.
2. Start with a 90% threshold On a new supplier. Check what the AI selects. If there are no complaints, reduce it to 85% for greater automation.
3. Don't disable rechecking, if the catalog is updated. 10 days is a good compromise.
4. Use a keyword filter for one-off tasks: “run only new products of this brand”, “process only large format”.
5. Open the window regularly (once a week) and process "not found" - these are new items that are waiting to be added to the catalog.
6. Before launching at a huge price (100,000+ lines) - Try it on parts first (filter by brand) to assess the quality and cost.
7. Auto-mode on several price lists — each price list is processed independently, in parallel. A task queue distributes the workload.
8. Don't run manual matching in parallel with AI There may be conflicts on the same price list. A yellow banner will warn you.
9. Check for links with the "AI-created" flag once a month per sample - this will give confidence in the quality of the AI work.
10. Export your connections to XLSX regularly. for backup and control.

Common setup errors

• Confidence threshold 50% — too low, it produces errors. Minimum 70.
• Number of candidates = 3 on a diverse catalog — AI may not see the correct product. Use 10+.
• Number of days to recheck = 1 — too often, wasting tokens. Minimum 5, usually 10–14.
• Auto-mode is disabled on regular price lists — you lose the whole point of automation.

🎉 Conclusion

AI-based product matching in Elbuz PIM is not just another feature, but a paradigm shift in how price lists are managed.

What previously required a team of three managers and weeks of work can now be accomplished in minutes at the click of a button. With quality control, decision transparency, error protection, and no double-spending on tokens.

You get not just a tool, but system AI handles routine tasks, the operator makes expert decisions, the log ensures everything is under control, and reports provide business data. Linking the price list to the master catalog ceases to be a chore and becomes a background process that requires no attention.

Start now. Import any price list, open the AI matching window, and click "Run." In just a few minutes, you'll understand why this system is already transforming the work of dozens of e-commerce stores on the Elbuz platform.

👉 If you want automatic operation right away, enable "Automatic mode" in the pricing settings. Elbuz and AI will then do everything for you. All you have to do is watch your data become cleaner and your managers work faster.

AI Matching: Connections Make Connections Easier.