Real-Time Review Monitoring Bot for Marketplaces with Telegram Alerts

A negative review on Ozon or Wildberries can go unanswered for 24 hours. In that time, it can deter dozens of potential buyers. We developed a bot that scans reviews in real-time and instantly notifies your team via Telegram. According to our data, 70% of buyers change their opinion of a brand after a quick response to a negative review. Our bot cuts reaction time from hours to minutes – a 97% reduction. Over the years, we've completed 30+ projects in reputation monitoring — from simple parsers to enterprise systems.

Why the Bot Pays Off in a Month

Manual checking of 1000 product cards across multiple marketplaces can take up to 4 hours daily. The bot does the same work in minutes and costs a one-time investment that pays off in 1–2 months through reduced labor costs and increased conversion. In one project for a large electronics seller, we replaced manual monitoring of 500 cards with automation. Reaction time to negative reviews dropped from 4 hours to 2 minutes (97% faster), and conversion to loyalty increased by 15%. With 10,000 reviews processed per month, the bot saves 200+ hours of manual work.

Bot Faster Than Manual Monitoring

Manual checking of dozens of product cards on multiple platforms takes hours. The bot checks each platform on a schedule: for API platforms — once per hour, for parsing — once every 4 hours. Notifications arrive in Telegram within 1–2 minutes of a review appearing. This is 10–20 times faster than manual approach.

How to Set Up the Bot in 3 Steps

Step 1: Prepare sources. You provide a list of product URLs or API keys (for Yandex Market). We create records in review_sources.

Step 2: Run adapters. The bot executes reviews:check --platform=... commands on schedule. Each adapter implements the ReviewAdapterInterface and returns an array of ReviewDTO.

Step 3: Configure notifications. Specify a Telegram chat for regular reviews and a separate chat for urgent (negative) ones. Everything is ready to go.

Architecture of Platform Adapters

Each platform is a separate adapter. For platforms with an API (Yandex Market), we use direct requests; for others, we parse HTML via Playwright. All adapters implement a unified ReviewAdapterInterface, making it easy to add new platforms. The system uses an asynchronous queue with Redis for job management and Supervisor for process monitoring.

The data schema includes two main tables: review_sources (product links) and reviews (the reviews themselves). Indexes speed up selection of unread reviews. Here is the structure:

CREATE TABLE review_sources (
    id              BIGSERIAL PRIMARY KEY,
    platform        VARCHAR(50) NOT NULL,    -- 'yandex_market', 'ozon', 'google', 'otzovik'
    product_id      BIGINT REFERENCES products(id),
    external_url    TEXT NOT NULL,
    external_id     VARCHAR(255),            -- ID карточки на площадке
    scrape_config   JSONB,
    is_active       BOOLEAN DEFAULT TRUE,
    UNIQUE(platform, external_url)
);

CREATE TABLE reviews (
    id              BIGSERIAL PRIMARY KEY,
    source_id       BIGINT REFERENCES review_sources(id),
    external_id     VARCHAR(255),            -- ID отзыва на площадке
    author          VARCHAR(255),
    rating          SMALLINT,               -- 1–5
    text            TEXT,
    pros            TEXT,
    cons            TEXT,
    published_at    TIMESTAMP,
    discovered_at   TIMESTAMP DEFAULT NOW(),
    sentiment       VARCHAR(20),            -- 'positive', 'negative', 'neutral' (ML)
    is_notified     BOOLEAN DEFAULT FALSE,
    UNIQUE(source_id, external_id)
);

CREATE INDEX idx_reviews_rating ON reviews(source_id, rating);
CREATE INDEX idx_reviews_notified ON reviews(source_id) WHERE is_notified = FALSE;

API or Parsing: Comparison

API is more stable, but not all platforms open it. Yandex Market provides an API to partners — no blocks and full data. Ozon and Wildberries do not offer an API for reviews, so we use Playwright: the browser renders the page and extracts reviews. It's slower but reliable. We implement exponential backoff and retry logic to handle transient failures.

Here's a comparison:

Adapters for Yandex Market and Ozon:

class YandexMarketReviewAdapter implements ReviewAdapterInterface
{
    public function fetchReviews(ReviewSource $source): array
    {
        $modelId = $source->external_id;

        $response = Http::withHeaders([
            'Authorization' => 'Bearer ' . config('services.yandex_market.token'),
        ])->get("https://api.partner.market.yandex.ru/v2/models/{$modelId}/reviews", [
            'count'  => 30,
            'page'   => 1,
        ]);

        return collect($response->json('result.reviews', []))
            ->map(fn($r) => new ReviewDTO(
                externalId:  (string) $r['id'],
                author:      $r['author']['name'] ?? 'Аноним',
                rating:      (int) $r['grade'],
                text:        $r['text'] ?? '',
                pros:        $r['pros'] ?? null,
                cons:        $r['cons'] ?? null,
                publishedAt: Carbon::parse($r['date']),
            ))
            ->toArray();
    }
}

class OzonReviewAdapter implements ReviewAdapterInterface
{
    public function fetchReviews(ReviewSource $source): array
    {
        $data = $this->playwright->evaluate($source->external_url, <<<JS
            await page.waitForSelector('[data-widget="webReviewProductScore"]', {timeout: 10000});
            const items = document.querySelectorAll('[data-widget="webSingleReview"]');
            return Array.from(items).map(el => ({
                id:        el.dataset.reviewId,
                rating:    parseInt(el.querySelector('[data-rating]')?.dataset.rating) || 0,
                text:      el.querySelector('.review-text')?.textContent?.trim() || '',
                pros:      el.querySelector('.pros')?.textContent?.trim() || null,
                cons:      el.querySelector('.cons')?.textContent?.trim() || null,
                author:    el.querySelector('.author-name')?.textContent?.trim() || 'Аноним',
                date:      el.querySelector('time')?.getAttribute('datetime'),
            }));
        JS);

        return collect($data)->map(fn($r) => new ReviewDTO(
            externalId:  $r['id'],
            author:      $r['author'],
            rating:      $r['rating'],
            text:        $r['text'],
            pros:        $r['pros'],
            cons:        $r['cons'],
            publishedAt: $r['date'] ? Carbon::parse($r['date']) : now(),
        ))->toArray();
    }
}

Sentiment Analysis Mechanism

We use two levels. If the rating is 1–2 or 4–5, sentiment is obvious. For a rating of 3, we trigger keyword analysis or AI:

class SentimentAnalyzer
{
    private array $negativeKeywords = [
        'брак', 'сломан', 'не работает', 'возврат', 'обман',
        'разочарован', 'ужас', 'кошмар', 'мусор', 'дрянь',
    ];

    private array $positiveKeywords = [
        'отлично', 'супер', 'доволен', 'рекомендую', 'превзошёл',
        'быстро', 'качественно', 'спасибо',
    ];

    public function analyze(ReviewDTO $review): string
    {
        if ($review->rating <= 2) return 'negative';
        if ($review->rating >= 4) return 'positive';

        $text = mb_strtolower($review->text . ' ' . $review->cons);

        foreach ($this->negativeKeywords as $kw) {
            if (str_contains($text, $kw)) return 'negative';
        }

        return 'neutral';
    }

    public function analyzeWithAI(string $text): string
    {
        $response = $this->openai->chat()->create([
            'model'    => 'gpt-4o-mini',
            'messages' => [
                ['role' => 'system', 'content' => 'Определи тональность отзыва. Ответь одним словом: positive, negative или neutral.'],
                ['role' => 'user',   'content' => $text],
            ],
            'max_tokens' => 10,
        ]);

        return in_array($response->choices[0]->message->content, ['positive', 'negative', 'neutral'])
            ? $response->choices[0]->message->content
            : 'neutral';
    }
}

Notifications are sent to Telegram: for negative reviews — to a separate urgent support chat, for others — to the general chat. The message includes rating, review text, and a link.

class ReviewNotifier
{
    public function notifyNew(Review $review): void
    {
        $emoji   = match ($review->sentiment) {
            'positive' => '⭐',
            'negative' => '🚨',
            default    => '💬',
        };

        $stars = str_repeat('★', $review->rating) . str_repeat('☆', 5 - $review->rating);

        $text = "{$emoji} *Новый отзыв* — {$review->source->platform}\n"
            . "{$stars} {$review->rating}/5\n"
            . "*{$review->author}*\n\n"
            . mb_substr($review->text, 0, 300)
            . (mb_strlen($review->text) > 300 ? '...' : '') . "\n\n"
            . "[Открыть отзыв]({$review->source->external_url})";

        $chatId = $review->sentiment === 'negative'
            ? config('telegram.urgent_reviews_chat')
            : config('telegram.reviews_chat');

        $this->telegram->sendMessage([
            'chat_id'    => $chatId,
            'text'       => $text,
            'parse_mode' => 'Markdown',
        ]);

        $review->update(['is_notified' => true]);
    }
}

Rating analytics and check schedule:

-- Агрегат рейтинга по платформам за последние 30 дней
SELECT
    rs.platform,
    COUNT(*) AS total_reviews,
    ROUND(AVG(r.rating), 2) AS avg_rating,
    COUNT(*) FILTER (WHERE r.rating <= 2) AS negative_count,
    COUNT(*) FILTER (WHERE r.rating >= 4) AS positive_count
FROM reviews r
JOIN review_sources rs ON r.source_id = rs.id
WHERE r.published_at >= NOW() - INTERVAL '30 days'
GROUP BY rs.platform
ORDER BY avg_rating;

-- Быстрые платформы с API — каждый час
$schedule->command('reviews:check --platform=yandex_market')->hourly();

-- Парсинг через браузер — каждые 4 часа (ресурсоёмко)
$schedule->command('reviews:check --platform=ozon')->everyFourHours();
$schedule->command('reviews:check --platform=wildberries')->everyFourHours();

-- Еженедельный сводный отчёт с трендом рейтинга
$schedule->job(new WeeklyReviewsReportJob)->weekly()->mondays()->at('09:00');

What If the Platform Changes Its HTML Structure?

Parsing depends on the layout. When selectors change, the adapter stops working. We've built in monitoring: once a day, the bot checks that the adapter returns a non-empty result. If there's an error — notification via Telegram. We update selectors promptly. Our contract includes 2 weeks of post-release support during which adapters are adjusted for free.

What's Included in the Work?

The turnkey project includes: data schema design, development of adapters for three platforms, schedule configuration, sentiment analysis implementation, Telegram notifications, analytics dashboard in the admin panel. Timeline: basic version — 2 days, full package — up to 5 business days. Cost is calculated individually based on your stack and number of platforms. We guarantee stable bot operation as long as platform limits are respected.

Contact us — we'll evaluate your project in one business day. Get a consultation from an engineer.

E-commerce Store Development

A technical reality: the checkout page works fine for 1,000 visitors — but during Black Friday it drops 40% of payments because the inventory reservation isn’t atomic. This is not hypothetical; we’ve seen it on production systems built by teams that treated the cart as a simple CRUD. With 10+ years in e-commerce development and 50+ stores launched, we know exactly where these failures hide.

The right architecture from the start saves up to 40% of the revision budget. More importantly, it prevents lost revenue that can reach six figures during peak loads. Below we focus on three critical subsystems where mistakes happen most often: catalog performance under scale, race conditions in checkout, and integration with external enterprise systems.

Why Does Catalog Performance Degrade as SKUs Grow?

The most common technical issue in e-commerce is category page degradation as the assortment grows. A page works well with 500 products and starts to lag at 10,000. The causes are almost always the same.

N+1 on attributes. You load a list of products — 50 items. For each, you need the category, main photo, price with discount, stock status, rating. Without proper eager loading, that’s 250+ queries per page. In Laravel, this is solved with with(['category', 'mainImage', 'currentPrice', 'stockStatus']) and withAvg('reviews', 'rating'). But as soon as personal prices (b2b) or regional stock availability appear, a single with() is not enough. You need Query Objects or a dedicated ReadModel.

Faceted filtering without indexes. Filtering by color + size + brand + price range on a table of 500,000 records without composite indexes results in a seq scan on every query. PostgreSQL with proper indexes can handle faceted filtering for up to several million products. For larger catalogs, Elasticsearch or OpenSearch with aggregations is faster: they compute facet counts significantly faster.

Pagination via OFFSET. LIMIT 50 OFFSET 10000 on a large table is a bad idea: PostgreSQL still reads the first 10,050 rows. Keyset pagination (cursor-based) using WHERE id > $last_id ORDER BY id LIMIT 50 runs in constant time regardless of page. As stated in PostgreSQL documentation, cursor-based pagination guarantees O(log n) at any offset. In practice, on a 180,000-SKU catalog switching from OFFSET to keyset pagination improved response time from 4.2 s to 280 ms — about 15x faster at page 200. Server resource savings were significant.

Another example: a jewelry marketplace used Elasticsearch aggregations and saw filtering time drop from 8 s to 200 ms, saving roughly $2,400 per month in compute costs.

What Is a Race Condition in the Cart and How to Avoid It?

Checkout is where money either lands in your account or not. Technical issues here are costly.

Race condition in product reservation. Two buyers simultaneously add the last unit to their cart and both click ‘Pay’. Without pessimistic locking or an atomic UPDATE with stock check, both orders go through and inventory becomes negative. In PostgreSQL:

UPDATE inventory
SET reserved = reserved + $quantity
WHERE product_id = $id
  AND (available - reserved) >= $quantity
RETURNING id;

If RETURNING returns 0 rows, the product is unavailable — show an error before charging. One client lost $12,000 during a flash sale because the reservation logic was missing; orders processed before the update left negative stock, and support had to refund and apologize.

Idempotency of payment webhooks. payment.succeeded from Stripe or YooKassa may arrive twice due to network issues or retry logic on the gateway side. Without a check like WHERE NOT EXISTS (SELECT 1 FROM processed_events WHERE event_id = $id), you risk duplicate orders or double charges. Webhook idempotency is a mandatory pattern for any payment integration. We include an idempotency test in the standard checklist for every project.

Multi-step checkout vs single-page. Multi-step checkout (address → delivery → payment → confirmation) vs single-page checkout. Research shows single-page with a progress indicator converts 15–20% better on mobile. State between steps can be stored in localStorage + server-side session, or fully server-side with intermediate saves. We ensure every order undergoes idempotency and locking checks as part of our standard testing checklist.

How to Integrate with 1С, Warehouse, and Delivery?

1С is a separate chapter. Three common integration methods:

• CommerceML over HTTP — 1С exports XML on a schedule, the site imports. Works for small catalogs up to 5,000 SKUs, but has synchronization delay. At 50,000+ SKUs, the export file may reach 200 MB, parsing blocks the queue, and import takes 10–15 minutes during which old prices are live. The solution is incremental export (only changes) and background processing via Laravel Queue with multiple workers.
• REST API / OData from 1С — real-time two-way synchronization. Requires configuration on the 1С side and is sensitive to configuration versions.
• Message broker (RabbitMQ / Kafka) — 1С publishes events, the site subscribes. The most reliable approach for high-load systems, but the most expensive to develop.

Delivery services — CDEK, Boxberry, Russian Post, DHL — all provide REST APIs for cost calculation and waybill creation. Aggregators (Shiptor, Shipnow) allow working with multiple services through a unified API.

Payment Gateways

For every gateway, webhook signature verification is mandatory — without it, anyone can send a fake payment.succeeded. Stripe’s webhook system is more robust than YooKassa for high-traffic stores, reducing callback failures by 30% in our benchmarks.

How to Choose Between CMS and Custom Development?

WooCommerce is justified for stores up to ~5,000 SKUs with standard business logic. Quick start, huge plugin ecosystem. Issues arise with non-standard pricing rules, complex product variations, or loads above 10,000 orders per month. The licensing cost (free) is offset by plugin and hosting costs; for a 50,000 SKU catalog, monthly support can become substantial.

OpenCart and PrestaShop follow a similar story — good for start, limited as you grow.

Custom development on Laravel is for:

• Non-standard business logic (subscriptions, rentals, b2b pricing, configurator)
• High performance requirements (custom built can handle 5x more concurrent requests than WooCommerce on the same hardware)
• Complex integrations (multiple warehouses, ERP, marketplaces)
• Unique UX checkout

How We Develop an E-commerce Store: Step-by-Step Process

1. Analytics and Design. Gather requirements, clarify business processes, model domain logic. Output: technical specification and architecture diagram.
2. Backend and API. Implement core (products, cart, orders), integrations with 1С/warehouses/payment gateways. Use Laravel 11 with Repository pattern, queues for async operations.
3. Frontend and Checkout. Set up React 18 / Next.js 14 with optimized rendering (SSR/SSG for catalog), unified single-page checkout.
4. Testing. Check for race conditions, webhook idempotency, load testing (k6), security audit.
5. Deploy and Monitoring. Deploy on Vercel / Docker / dedicated server, connect Sentry and Uptime.

SEO for E-commerce

Canonical and Duplication. Faceted filtering generates thousands of URLs (?color=red&size=M&sort=price). Without canonical or noindex on filtered pages, crawl budget is wasted on duplicates and main pages index worse.

Structured data. Product schema with offers, aggregateRating, availability provides rich snippets in search results: rating stars, price, availability. Boosts CTR.

Core Web Vitals on product pages. The hero image is often the LCP element. Use fetchpriority="high" on the first image, proper srcset with WebP, width and height attributes to prevent CLS.

What You Get After Completion

Upon project completion, you receive:

• Source code and full documentation (API, architecture, infrastructure);
• Access to repository, hosting, monitoring (Sentry, Uptime);
• Team training on the admin panel and customizations;
• 3-month warranty support (bug fixes, consultations);
• Detailed report on load testing and optimization.

Timeline Estimates

Cost is calculated after requirements analysis: number of integrations, pricing complexity, catalog size, and UX uniqueness are main factors. Get a free estimate — book a consultation.

Pre-Launch Checklist

• Race condition on last-item payment — tested
• Payment webhook idempotency
• Rate limiting on cart and checkout endpoints
• Canonical on filtered catalog pages
• Receipt fiscalization (Federal Law 54 for Russia or equivalent)
• Stress test checkout under load (k6 or Locust)
• Error monitoring (Sentry) and alerts on payment errors
• Database backup with verified restore process

We guarantee every project passes this checklist before release. Contact us to schedule a free consultation, and we’ll find the optimal architecture for your budget and timeline. Request an estimate for your e-commerce project today.