What we learned running our own websites

The ground: sites that sell, measured closely

All our sites follow the same set-up: fast static pages, one-click online payment, measurement through Search Console, Bing Webmaster Tools and Google Analytics 4 with purchase events. 31 of them send e-commerce events to Google Analytics.

Over 90 days (20 July 2026 to 17 October 2026), these sites earned 479,265 impressions and 4,715 clicks in Google, a click-through rate of 0.98%. 25,409 pages appeared at least once in the results; only 11% of them got a click. That is the first lesson: an indexed page is not a page that works.

Lesson 1: mass publishing costs more than it earns

During summer 2026, 10 of our sites lost 90% to 99% of their impressions. Our investigation of 22 September 2026 found no technical fault. Every drop followed a mass release of generated pages or assembly-line publishing, with template pages 56% to 73% similar to each other. The median position of entire sites went from about 24 to 70 to 80.

The figures for the group of 13 sites launched in July show it month after month: 3,337 impressions a day in July, 2,726 impressions a day in August, 1,691 impressions a day in September, 1,848 impressions a day in October. Of the 15 sites with more than 70 days of data, 7 are below half of their best 28-day period and 5 below 10%.

The sites that held up have one thing in common: every page carries its own data point (a calculation, a measurement, an official source) that cannot be found elsewhere. Google says so in its guidelines: it looks for helpful content made for people, and its spam policies target mass page production with no added value.

What we have done since: at most one new page a day per site, on a real search found in Search Console, with a data point or source of its own on every page. Better to publish less than to force it.

Lesson 2: pruning loses nothing, merging can bring traffic back

On a van comparison site, we removed 8,194 pages with a 410 code on 23 September 2026 (those with no impressions in 90 days), then merged 378 pages with 301 redirects on 28 September 2026. 810 indexable pages remain.

The result, over comparable periods: 104 impressions a day before, 98 after; average position 69.4 before, 64.5 after; 663 queries before, 665 after. Removing most of the addresses cost nothing. But nothing came back either: pruning cleans up, it is not enough to win traffic back.

Three other sites had merged their page variants with 301 redirects on 22 and 23 September, just before Google’s spam update of 24 September. All three came back: a pet site with a shop, from 1,955 to 147,302 impressions in 28 days (position 56.2 → 11.8); an automation templates site, from 738 to 9,546 impressions in 28 days (position 22.6 → 8.5); a sales page templates site, from 535 to 4,973 impressions in 28 days (position 62.1 → 13.0). The cause is not proved, and other sites pruned the same way stayed at rock bottom. We take a cautious rule from it: merge the pages that answer the same intent, keeping their content, rather than multiplying them.

Lesson 3: Bing and AI assistants already sell

From 1 October 2026 to 19 October 2026, 52% of visits from a search engine on our selling sites came through Bing and the engines that use its index (Ecosia, DuckDuckGo, Yahoo), against 47% through Google. On 4 sites, Bing beats Google. On two guide sites that Google hardly indexes any more, every search engine visit comes from this family.

Over the same period, 13% of measured visits came from an AI assistant, and ChatGPT accounts for 92% of them. The sample is small (Google Analytics only counts visitors who accept measurement), but the order of magnitude is clear: a site missing from Bing and unreadable for AI leaves buyers to others.

What we do on every site: Bing Webmaster Tools, new pages sent through IndexNow, AI crawlers allowed, llms.txt, a first sentence that answers, structured data that matches the page.

Lesson 4: every step before payment costs sales

We tested several journeys on our guide sites and our shops. What stayed are the decisions that remove a step:

• 4 October 2026: One click to the payment page: price visible, no intermediate page, no account.
• 4 October 2026: llms.txt, AI crawlers allowed, a first sentence that answers, structured data that matches the page, IndexNow, on every site.
• 5 October 2026: Trust signals from true facts only: one line under the button, a three-step “how it works” block; no counter, no urgency, no fake badge.
• 5 October 2026: GA4 add_to_cart, begin_checkout and purchase events and a consent banner on every site that sells.
• 10 October 2026: Simplification: a menu of 5 items at most, a single buy label, no more promo banner or promo code without an end date.
• 12 October 2026: Bundles from €49 to €149 that group small products, rather than sales at €5 to €12.
• 12 October 2026: Repeat purchase in one button (card kept by the payment provider with consent), Google Pay enabled.
• 17 October 2026: Every page three clicks or fewer from the home page through real HTML links, checked every week by an in-house crawler.
• 17 October 2026: Optional email before payment (“to find your purchase again”), a basket only from 3 products, pay first, personalise afterwards.

As an order of magnitude only (small sample, 351 measured sessions), there are about 1.4% add-to-carts and 1.1% purchases per session. These are not benchmarks to aim for: what counts is how your own rate changes after each change.

Lesson 5: check the margin before launching

Dropshipping promises a shop without stock. Our test of 8 October 2026 compared supplier price and retail price, product by product: of 38 expensive products held in European stock, none reached the 1.8 multiplier we were aiming for between cost and selling price. Of 14 products in US stock, the brand itself sold at the supplier’s price.

So we changed our approach: products whose value we control (guides, plans, templates), grouped into bundles, and physical items chosen only when the margin is real.

Lesson 6: mobile first, long searches first

61% of measured visits happen on mobile, as do 55% of clicks from Google. A buying page is therefore designed for a thumb first: a button visible without scrolling, the price at the top, a short form.

On the search side, Google hides the query behind 93% of our sites’ clicks: these are rare searches, too infrequent to be shown. Among visible queries, those of four words or more make up 39% of clicks and 47% of impressions. The traffic that sells is made of thousands of precise questions.

What we apply for our clients

These lessons have become our working method, in this order:

1. 1
   How Google and Bing read the site: Search Console, Bing Webmaster Tools, sitemap, IndexNow.
2. 2
   Measuring the journey: GA4 purchase or enquiry events, with consent handled properly.
3. 3
   A shorter buying or enquiry journey: one click to the payment, or a three-field form.
4. 4
   Existing pages rewritten: an answer from the first sentence, verifiable facts, real proof.
5. 5
   Pruning and merging the pages that bring nothing, internal linking at three clicks at most.
6. 6
   New pages at the pace of real searches, never in bulk.