Do LLM training pages actually work?
Yes. In a controlled eight-week test, the category that got training pages was crawled about 17% more than a comparable one held back as a baseline. One product category got prerendered LLM training pages. A matching category was held out of the rollout and used as a baseline. Eight weeks later the treated category was being crawled 17% more than the holdout. Crawling rose across the whole site over the same weeks, so without the holdout there would be no way to tell our effect apart from the background. The page we changed is an ordinary category listing page, so the same build applies to ecommerce pages on Shopify, Magento, Adobe Commerce, Salesforce Commerce Cloud, and anything else that can render HTML on the server. A training crawl is how a model learns your brand by heart. Get it right and the assistant names your product from what it already knows: no web fetch, no citation, no link to click, just the specifications and comparisons it read once and kept. That only happens if a training crawler could read the page to begin with, and it is the one step in the chain a brand can directly control and directly measure. What follows is the page we shipped, the eight weeks after it went live, and the three tests we ran to try to break the result.
- AI training crawls ran 17% above the untouched control. Measured over the 8 weeks after launch. The result holds through a placebo test on 46 invented launch dates, through dropping any one page from the analysis, and through changes to the analysis window.
- A comparable category ran alongside as a holdout, and every number is measured against it. Site-wide crawling rose sharply over the same window, so the holdout is what separates the effect from the background. It also carried the data-quality checks: when training crawls read zero for seven straight days across every category at once, the holdout included, that identified a logging outage rather than crawler behaviour, and those days were dropped.
- Both categories more than doubled. The training pages account for the 22% gap between them. Raw training crawls rose 2.7x on the treated category and 2.2x on the holdout that received nothing at all, so most of that growth was industry-wide and would have arrived anyway. Indexed against their own pre-launch averages, the two lines end the window 22% apart. Averaged across the full eight weeks the gap settles at 17%.
-
In-stock pages gained 43% of training crawls. Out-of-stock pages lost 22%.
In-stock pages (10): 5,090 → 7,270 training crawls (+43%)Out-of-stock pages (3): 870 → 680 (−22%)Every out-of-stock page lost crawl attention. This comes from 13 pages and we did not predict it in advance, so it is a lead we would want to test properly rather than a finding.
- Once a model has read your pages, it can recommend them without looking them up. A training crawl is what puts your product facts into the model itself, so an assistant can answer from memory rather than going out to the web. Getting crawled is not the same as getting recommended. What it changes is what the model knows about your range by default, before anyone asks it anything, and it is the part of the chain a brand can change on purpose and check the next morning. Crawls are the only thing we measured here.
One category got training pages. A comparable one did not.
Our client shipped prerendered LLM training pages for one product category in May 2026. A second category of comparable size was held out of the rollout so it could be used as a baseline. Everything in this report is measured against that holdout.
A "training page" here is a prerendered, server-rendered page built so that an AI crawler can read the whole product story without executing JavaScript: specifications, comparisons, and plain-language answers, with body-content links out to each product page. The question is whether shipping them changes anything a crawler does.
Answering that is harder than it looks, because AI crawling is rising everywhere. Ship training pages in May, watch crawls go up in June, and you have learned very little. You need to know what would have happened anyway. The holdout category supplies that: it is the closest available estimate of what the treated category would have done if nothing had been shipped.
Categories are renamed after planets. The product pages inside the treated category are renamed after moons. The numbers are exactly as measured. Our client and their vertical are not disclosed.
The metric is AI training crawls: hits from AI training crawlers on the product pages that each category's listing page links to from its body content. Navigation, header, and footer links are excluded, so we are counting crawls on the pages the training page actually points at.
The unit of observation is the day, not the hit. One crawler session can emit hundreds of requests in a burst, so treating 17,030 hits as 17,030 independent draws would make almost anything look significant. There are 118 days on the charts below, and 57 days on each side of launch inside the statistical windows.
Once a model has read your pages, it can recommend them without looking them up.
A training crawl moves your product facts into the model itself, so an assistant can answer from memory instead of going out to the web. That is the metric this study moves.
Here is what that looks like in practice. Someone asks an assistant which one of these they should get. It answers by naming your product, describing what it does and who it suits, and comparing it against the alternative, all from what it already learned. There is no fetch, no citation, and no source link. The model knows your range the way it knows anything else, and that knowledge came from pages a training crawler read and absorbed months earlier.
Live retrieval, where an assistant fetches a page mid-answer and quotes it, is the other route in. It matters, and it starts in the same place: a crawler asking your server for a page and being handed something it can read.
Hand it a JavaScript shell and you are in neither. The model does not learn what your product does, the retrieval layer has nothing to pull, and when a shopper asks an assistant to compare your range against a competitor's, the assistant answers from whatever it does have. That is usually a marketplace listing, a review site, or the competitor.
That is the honest version of why this number matters, and it is also the reason we are careful about what we claim from it. This study proves the training pages got the pages read more often. It does not prove what any model did with what it read. We are not going to pretend otherwise, and a report that jumps straight from a crawl chart to a revenue promise is a report you should not trust.
What makes the crawl worth chasing anyway is that it is the one link in the chain you can act on directly. You cannot make a model cite you. You can decide what your server hands a crawler, and you can count what happens next.
Roughly 3,000 words of prerendered content sit below the product grid.
"LLM training page" is a vague phrase, so we pulled the live page apart and counted what was on it.
The listing page keeps its normal product grid at the top. Everything below the grid is the training content: prerendered HTML a crawler can read without executing JavaScript. It runs to roughly 3,000 words, about 77% of the body copy on the page.
The structure repeats itself deliberately. Every product deep-dive has the same shape: a "what makes this one different" section split into six separately-headed reasons, followed by a specification table. A model reading it has little to infer, because each answer already sits as a self-contained chunk under a heading that names the question it answers.
A great deal of AI visibility advice is really schema advice, so we left the schema off. In-copy links to the products were left off for the same reason. Whatever moved, moved because the content was in the HTML, organised under headings and into tables, and readable without a browser. That gives us a clean floor to build on: the next test adds the internal links, the one after that adds the structured data, and each time we can measure what the addition is worth on its own rather than shipping everything at once and guessing which part did the work.
One more thing held back on purpose: only 6 of the 13 linked product pages got a deep-dive block of their own. The other seven are linked from the grid and nothing more. That was not an oversight. It leaves seven pages that received the benefit of sitting on an enriched page without any copy of their own, which turns out to be one of the more useful comparisons in the study.
Training crawls rose 2.7x on the treated category and 2.2x on the holdout. The gap between them is 22%.
Each line is indexed so its own pre-launch average is 100, so both start level. Both more than double over the window, including the holdout, which received nothing. Only the 22% they finish apart can be credited to the training pages.
In raw weekly numbers the treated category went from 630 crawls a week to 1,690, a 2.7x rise. Over the same weeks the untouched control went from 1,670 to 3,650, a 2.2x rise, with no training pages at all.
AI crawling is climbing on nearly every site at the moment, so almost any change you ship will be followed by a rise in crawls. A rise on its own therefore carries very little information. What carries information is growing faster than a comparable category that was left alone.
The treated category outgrew the control by 17%.
The same data with the rising tide divided out. The treated category is divided by the control, then indexed so the pre-launch average is 100. Above 100 means it grew faster than the rest of the site.
Before launch the line wanders around 100 with no trend, which is what a valid control looks like: the two categories were moving together. After launch it lifts and stays lifted. The line's own post-launch average is 114. The formal estimate, which is a log-scale regression rather than an average of ratios, is +16.8% with a 95% confidence interval of +7.8% to +26.4%. Two different estimators, both landing inside the same band.
The blue line and the shaded band are different objects. The line is a daily ratio and any single day of it means very little. The band is the estimated average lift across the whole post-launch window, and its height is the uncertainty around that average. The band is the finding. The last point on the line is not.
How we validated the finding: three tests, and it held through all three.
One test returning a good number is not worth much on its own. These are the three we ran against it.
46 invented launch dates produced nothing.
We re-ran the whole analysis against 46 launch dates we made up. A method that finds effects on those is finding them in noise. None of the 46 produced an effect as large as the real one.
No single page is carrying the result.
We dropped each of the 13 pages in turn and re-ran. The result held every time, ranging from +10% to +19%. No single page is carrying the finding.
The estimate is stable across settings.
Different window lengths and autocorrelation lags move the estimate by a few points and never change the conclusion.
One pattern is worth showing on its own. Crawling runs on a strong weekly cycle, and the change after launch is very uneven across the week. Any comparison that does not line up like-for-like on the day of the week is partly measuring the calendar.
Saturday fell. Every other day rose, and Thursday and Sunday rose most.
One correction, and it is worth spelling out. Any result like this could in principle be a coincidence, so the first thing we do is ask how big a coincidence it would have to be. Our first pass said the odds of getting this by luck alone were about 1 in 3,000.
The placebo test told us that number was too flattering. When we pointed the same maths at dates where nothing had happened, it still came back claiming a result far more often than it should have, so the method is more easily impressed than it lets on. We marked our own confidence down to match. A fair statement of the odds is closer to 1 in 50. That is still a real result, just a more modest claim than the first one.
The odds are not the number to hang on to. The range is. Our best estimate of the lift is 17%, and the bracket around it runs from 8% to 26%. That is the sentence we would defend: the training pages produced somewhere between 8% and 26% more crawling, and most likely around 17%.
The extra crawling landed on the products that were in stock.
The category average hides the shape of this. Broken out to the 13 product pages, the extra crawling lands very unevenly.
Pooling the pages before smoothing is what makes this readable. A single product page gets a handful of crawls a day, and one line of that is mostly jitter. Added together, the ten in-stock pages hold flat through the spring and then step up after launch. The three out-of-stock pages do something different, and it is not quite the story we expected.
The out-of-stock line was already sliding before the training pages existed. It comes into the launch date low and drops further afterwards. We cannot hand the whole of that decline to the training pages, and we are not going to. Some of it started earlier, for reasons this study did not measure.
Page by page the noise is obvious, which is the point of showing it. Individual panels wander a long way from their own baseline in both directions before anything was shipped. What survives that noise is the tally at the end of the window.
And it is not only stock status. Remember that only six of the thirteen pages got a deep-dive block of their own. Sort the pages by what the training page actually gave them and a gradient appears.
Pages with a written block of their own gained most. Pages that were only linked from a page that got richer still gained, which is the more surprising half: a product does not need its own copy to benefit. Out-of-stock pages lost ground either way.
All three out-of-stock pages lost crawl attention across the window. Nine of the ten in-stock pages gained it. The exception, Pandora, is a low-value line item rather than a primary product, and it fell 32%.
That claim rests on 13 data points and we did not predict it in advance. No individual page here is significant on its own and we have attached no p-value to any of them. Part of the out-of-stock decline also predates the launch, as the pooled chart shows. It is a lead worth testing properly on another category, not a result.
The finding covers training crawls. It does not stretch further than that.
What this design measures well, and where it runs out.
Training crawls are the metric this study can carry. The daily counts are large enough, the holdout is clean, and the effect survives every test we ran at it. Everything in this report is a statement about how often AI crawlers fetched these pages, and nothing more than that.
Downstream behaviour is a different question. Referrals from AI assistants came in at +9% against the holdout, but the smallest change this design could reliably detect is around ±72%, so that number is noise dressed up as a result. We are not going to report it as a finding. Measuring what happens after the crawl needs either far more traffic or far more time, and probably both.
One category is not a law. This is a single treated category measured against a single holdout over eight weeks. It is enough to say the mechanism worked here. It is not enough to say how large the effect would be in another catalog, in another vertical, or a year from now.
Building your first LLM training page: what to put on it, and how to know if it worked.
This is the page we measured, described as a build spec, plus the measurement setup we would insist on before anyone ships anything.
The page that produced a 17% lift is a category listing page with a long, well-organised block of prerendered HTML underneath the product grid. There is no trick in it. The steps below are what that page is made of, in the order we would build it again.
Put it on the category page, below the grid
Not on a new URL, and not hidden behind a tab or an accordion. The page you are enriching is the one that already links to every product in the range, and those product links are what the crawler follows next.
Do the boring technical SEO first, because none of this works without it
The page has to be reachable before it can be read. Confirm it returns a 200 and is not blocked in robots.txt for the AI user-agents you care about, that it is in the XML sitemap, that it is linked from somewhere real rather than orphaned, that the canonical points at itself, and that it is not noindex. Every number in this report depends on a crawler being able to reach the page, and none of the writing matters if it cannot.
Prerender it. This is the part that is not optional
The text has to be in the HTML that arrives from the server. If a crawler has to run JavaScript to see your content, assume it will not see your content. Fetch the page with JavaScript disabled and read what comes back. That is your training page.
Nothing here is specific to one platform. A category listing page is a category listing page, and every major ecommerce stack can render one on the server. What changes is what is likely to break it.
The test is the same everywhere and it takes a minute: request the page with JavaScript turned off and read what comes back. If the block is not in that HTML, it does not exist as far as this study is concerned.
One block per product you actually want sold, about 400 words
Six products got a block here, averaging about 400 words. Each one followed the same skeleton: a "what makes this one different" section, then a specification table. Repetition across products is a feature. It makes the whole range easy to parse and easy to compare.
Break the pitch into about six separately-headed reasons
Not a bullet list, and not one flowing paragraph. Each reason gets its own subheading and about 30 words underneath it. The heading should name the thing it answers, so the answer can be lifted out on its own and still make sense.
Give every product a specification table
Around 14 attribute rows: whatever a buyer would actually compare before choosing between two of your products. A table states the fact and the label for the fact in the same row, which is why it survives extraction so much better than the same information written as prose.
Add one comparison table across the whole range
Same attributes, every product, price included. This is the single most quotable object on the page: it is the answer to "which one should I get", which is the question people actually put to an assistant.
Finish with four or five real buying questions
Questions as headings, answers of about 100 words, and a small table inside the answer where a table helps. Write the questions your sales team gets asked, not the ones a keyword tool suggests.
Point it at products you can actually sell
Every out-of-stock page in this study lost crawl attention while the in-stock pages gained it. That split is unproven and part of the decline predates the launch, but acting on it costs nothing. Do not spend crawl budget advertising products nobody can buy.
Choose a holdout before you build anything
Pick a comparable category of similar size and traffic, and agree in writing that it does not get training pages this quarter. Without it you cannot tell your effect apart from the industry-wide rise in AI crawling, and you will end up reporting the rise as if you caused it.
Collect at least eight weeks of baseline first
You need enough history to show the two categories were moving together before you changed one of them. If they were already diverging, the comparison is worthless, and you can only discover that from the baseline.
Log AI crawler hits at the edge, per URL, per day
Server or CDN logs, not a JavaScript analytics tag: the crawlers you care about never run the tag. Keep the daily granularity. The day is the unit you will analyse, because one crawler session can fire hundreds of requests in a burst and counting hits would make almost anything look significant.
This is what WISLR AI Channel Analytics was built to do. It reads AI crawler traffic at the edge, per URL and per day, from your own logs rather than a third-party estimate, which is exactly the data every number in this report is built on. If you do not want to assemble the pipeline yourself, that is the shortcut.
Report the gap, never the raw growth
Your crawls will go up. So will the holdout's. The only number worth putting in front of ecommerce leadership is the difference between the two, and any report that shows you crawl growth without a comparison group is showing you a chart rather than a result.
Re-measure in a quarter
The strongest weeks of this study were the last two. That is either compounding as crawlers re-index the range, or it is noise, and eight weeks cannot tell the difference. Book the follow-up now, while the holdout still exists.
We can run this study on your catalog.
We build the training pages, hold a comparable category back as a baseline before anything ships, and measure what the pages did against it. That is the difference between reporting a 2.7x rise and reporting the 22% of it you actually caused.