Competitor monitoring usually collapses into one of two failure modes: doing it by feel, with someone manually checking a few pages every couple of weeks, or writing a scraper that hammers pages until the IP is blocked, leaving a pile of data nobody reads.
The way through is to narrow the question first: what change would make you do what? Only fields that can trigger an action are worth monitoring.
Step 1: define the fields
Don't "capture the competitor page". Make a table and keep only what changes a decision:
| Field | What a change means | Possible action |
|---|---|---|
| Price / promo price | The price band moved | Reprice, adjust bids, or do nothing |
| In stock | Supply changed | Raise ad budget, check your own stock |
| Main image / image count | Visual strategy changed | Review your own hero image |
| Title and key attributes | Positioning or keywords changed | Re-check your keyword coverage |
| Rating and review count | A rough demand signal | Observe; never a conclusion on its own |
| Variant count | Product line expanding | Add to sourcing review |
Fewer fields, less noise. For most teams the real list is price, availability and title.
Step 2: official channels first
Two lines govern collection: the platform's terms of service and the target site's robots rules. In order of preference:
- seller-side data and reports the platform already gives you;
- official or licensed pricing and market data services;
- low-frequency human sampling: one person records 10 key SKUs at a fixed time each week;
- restrained, low-frequency access to public pages, respecting robots directives and rate limits.
What not to do: high-frequency scraping, bypassing logins or verification, impersonation, collecting personal data, or copying competitor page content into your own listings. Trading account and legal risk for a few saved hours is a bad deal.
Start low-frequency and running, rather than designing a perfect system that never ships.
Step 3: let AI classify diffs, not draw conclusions
Most of what you collect is noise. AI is good at turning "what changed" into a summary a human can scan:
You are a competitor monitoring assistant. Input is two field snapshots of the same SKU at two times.
Tasks:
1. List only the fields that changed, with old value, new value and magnitude;
2. Group by type: price / stock / content / variants / reviews;
3. Flag changes needing human attention (price move beyond threshold, out of stock, core title term changed);
4. Do not speculate about competitor intent and do not recommend actions.
Data: ...
That last instruction matters. Ask a model why a competitor cut its price and you'll get a plausible-sounding guess every time. Classification and summarising go to the model; judgement stays with a person.
Step 4: write the response as a rule, not a reflex
Agree thresholds and actions in advance, so nobody instinctively matches every price cut:
| Situation | Pre-agreed action |
|---|---|
| Competitor price cut under X% | No action, log it |
| Cut over X% sustained more than N days | Check margin, then decide on repricing |
| Competitor out of stock | Raise ad budget on that term, check own stock |
| Core title term changed | Review your own keyword coverage |
| New variant added | Queue for sourcing review, don't follow immediately |
Set thresholds from your own margin, not someone else's playbook. Matching prices is the most over-used action available: entering a price war takes a click, leaving one takes months.
Step 5: control frequency and volume
- price and stock: daily is usually enough, unless the category is genuinely volatile;
- title, main image, variants: weekly;
- review counts: weekly or monthly, read as a trend rather than a point;
- monitor only the SKUs on your list, not a whole category.
Give the data a retention limit too. Three years of competitor price snapshots does nothing but take up space.
Four common misreads
- Treating a promotion as a permanent cut. Confirm it held for several days.
- Treating review growth as sales. Correlated, not equivalent; incentivised reviews distort the ratio.
- Watching one competitor. A single rival's move may just be their own stock problem.
- Treating the AI summary as the decision. It saves you page-flipping; it doesn't set your price.
What one snapshot row should hold
Whether the data comes from a report export, a weekly manual check or a page visit, keep every snapshot in the same shape so diffs are possible at all:
date, your_sku, competitor, competitor_url, price, promo_flag, in_stock,
title, image_count, variant_count, rating, review_count, sourcepromo_flag stops you mistaking a coupon or deal for a permanent cut. source records how the row was collected (report, manual, page), which matters when two rows disagree and when someone asks later whether collection stayed within the platform's terms.
Worked example: one SKU, one week
Take a fictional seller of an insulated steel water bottle as an example. They track three direct rivals on that one SKU, and their pre-agreed rule says a cut over 10% sustained for more than 3 days triggers a margin check.
The weekly diff summary from the model reads, in effect:
- Price: rival A down from 24.99 to 21.99 (−12%), flagged as beyond threshold;
- Stock: rival B out of stock since Wednesday, flagged;
- Content: rival C's title changed "insulated" to "vacuum insulated", flagged as a core title term change;
- Reviews: all three up slightly, not flagged.
What the person does with it: checks promo_flag on rival A and finds the lower price is a limited deal, so it's logged and re-checked in three days rather than matched. Checks their own stock before nudging ad budget on the term rival B dropped out of. Adds "vacuum insulated" to the keyword coverage review instead of rewriting the title that afternoon. Nothing gets repriced, and every decision traces back to a row in the rule table.
A prompt for the weekly roll-up
Once daily diffs are running, a second prompt turns a week of them into one page matched against your rules:
You are a competitor monitoring assistant. Input is this week's change log for [NUMBER] SKUs
in [CATEGORY], plus my pre-agreed response rules.
Tasks:
1. For each flagged change, state which rule row it matches, or "no rule";
2. List changes that match no rule in a separate section, so I can decide whether a rule is missing;
3. Count changes per competitor and per type (price / stock / content / variants / reviews);
4. Do not guess at competitor intent and do not suggest prices or actions.
Rules: [PASTE YOUR RULE TABLE]
Change log: [PASTE THIS WEEK'S DIFF SUMMARIES]The "no rule" section is the useful part. If the same unmatched change keeps turning up, write a rule for it; if it never leads to an action, stop collecting that field.
Implementation checklist
- three to five monitored fields, each tied to a defined action;
- collection method complies with platform terms and robots rules, at a restrained frequency;
- someone actually reads the diff summary at a fixed time;
- response thresholds fixed in advance, with every action and outcome logged;
- data retention limited;
- no competitor copy pasted into your own listings.
FAQ
How many competitors should I monitor per product? Enough that one rival's stock problem doesn't look like a market move, and few enough that someone reads the summary. In practice that's a handful of direct rivals per SKU: the listings you actually lose sales to, not everything in the category.
Is it legal to scrape competitor prices? It depends on the platform's terms, the site's robots rules and the law where you operate, which is why the order above puts your own seller reports and licensed data first. If you're unsure whether a collection method is allowed, treat that as a no until you've checked.
Should I hand this to an automatic repricer? Whatever tool ends up changing the price, the thresholds and your margin floor have to be written down first. A repricer that matches every cut is the price war from Step 4 running on autopilot.
Can AI tell me why a competitor changed its price? It will always offer a reason, and it has no way of knowing. Use it to sort and summarise changes; the "why" comes from what you can actually verify, such as a visible promotion or a stock-out.
For the research methodology side, pair this with Perplexity vs ChatGPT for market research. For your own listings, see optimising Amazon listings with AI.
The scheduled collection and alerting layer usually sits in an automation tool. On whether to run your own, see n8n or Make: should a small team self-host its automation.
Method compiled 2026-08. Before collecting data from third-party sites, confirm the target platform's terms of service and the legal requirements in your jurisdiction.
