How to spot an AI slop website: 7 tells (with real examples)

A slop website is not one that used AI. It is one that was published without anybody reading it first: the template stayed a template, the placeholders stayed in, and every heading came out of the same mold. Those leftovers are what you can spot.

To find out which tells still work in 2026, we went through the 30 most recent public reports on CrawlRanker, all scanned between October 16 and 19, 2026. Our AI slop detector read 17 of those pages as Mostly AI-generated, 8 as Mixed signals and 5 as Mostly human-written. Every example below comes from one of those reports, with the verdict and confidence it got. We don’t name the sites: the point is the pattern, and nearly every one of these pages also did something well. Generic wording is quoted exactly. Where an exact quote would point to a single site, the identifying words are swapped for a description in [italic brackets].

First: the old tells have stopped working

The 2023 way to spot slop was a word list: “delve,” “leverage,” “seamlessly,” “in today’s fast-paced world.” We still count those phrases on every scan. Of the 17 pages read as Mostly AI-generated, 15 contained none of them. The other two had one hit each: a single “robust” and a single “the world of.” Across all 30 pages, only 4 used any stock phrase at all.

The second classic tell, flat and even sentence rhythm, fared no better. Not one of the 30 reports rated a page’s rhythm as uniform. Of the Mostly AI-generated pages with enough prose to measure, 9 of 12 came out as varied, the label human writing usually gets.

Models have learned to avoid the words people mock. What they haven’t learned is what nobody asks them for: to check their own output against the page it lands on. The seven tells below are all versions of that.

1. Placeholder text nobody deleted

The fastest tell to check and the most damning. A generated draft marks the gaps it couldn’t fill, and slop ships with the gaps still marked:

Placeholder residue proves that nobody proofread, not that a model wrote the page. A hand-built WordPress site in the same batch still showed its theme’s demo street address and a “yourmail@…” email in the header, and it read as Mostly human-written. That is exactly the distinction slop is about: effort, not authorship.

Check it: search the page for square brackets, “lorem,” “example,” “your company” and phone numbers like 123 456.

2. The landing page every builder ships

Hero, three feature cards or a column of checkmarks, a numbered “how it works in 3 steps” block, an FAQ, a closing call to action. A templated-structure finding showed up on 16 of the 17 Mostly AI-generated pages, most often for this very sequence. The headings give it away, because they would fit any product. Four of them, from four different pages:

A template on its own isn’t slop; plenty of careful sites use one. It becomes a tell when nothing in the sections could only be on this site. The slop detector’s structure axis weighs the layout against the voice and specifics around it for that reason.

Check it: write down the section order. If it matches the list above and you could swap in a competitor’s name without changing a sentence, the template is doing the work.

3. Slogan headings in matched pairs

Generated marketing copy loves a balanced pair: “X, not Y,” “X. No Y.” or “built like A, priced like B.” One pair is a style. A page of them is a pattern:

The last one matters: the “not this, but that” rhythm survives translation, so it works as a tell in any language.

Check it: read the headings aloud in a row. If they sound like a stack of taglines, nobody wrote them for this page.

4. One keyword, repeated like a chant

Slop often comes from a pipeline told to rank for a phrase, and the phrase leaks into every paragraph whether it fits or not:

The report found that same keyword, ending in “Online free,” in almost every paragraph and heading, including in sentences where no human would say it. Old-fashioned human SEO does this too: the WordPress site from tell 1 repeats a “Best [specialty] in [city]” phrase several times and still read as Mostly human-written. Keyword chanting tells you the page was written for a crawler. On its own, it doesn’t tell you by what.

Check it: search the page for its own main keyword. More than one awkward, grammatically bolted-on hit per section is the tell.

5. Sentences about the page instead of facts

This is the subtlest tell and, in 2026, the most common one. The copy describes what the page does in abstract meta-language instead of saying anything you could check:

Each of these sentences is grammatical, plausible and empty. None contains a number, a name, a date or a place.

Check it: pick five sentences at random and try to underline one fact in each. Written pages describe the world. Slop describes itself.

6. Images nobody looked at

Media is where the missing review shows most plainly, because checking images takes the longest. A travel guide to a coastal village, read as Mostly AI-generated at 72% confidence, captioned six of its section images as AI concept renders.

To its credit, that page labels its renders honestly. The tell is not that a model made the pictures; it’s that the place the guide sells never appears in a real photo. Elsewhere in the batch, a birthday-page builder had 41 of its 54 images without alt text, and a checklist tool had no images or screenshots at all to show the product it described.

Check it: look for one photo that could only have been taken by this business, then inspect a few images for alt text. Generic renders plus blank alt text means nobody went through the media.

7. Copy-paste rows and leaked edit notes

When a page is generated in pieces, the seams show as identical blocks and, sometimes, as instructions that were meant for the model.

A motor maker’s article index, read as Mostly AI-generated at 60% confidence, repeated the same one-line summary word for word in all three rows of a table, and did the same in its “next step” column. An engineering calculator, at 66%, went further: its FAQ asked why the page had removed a particular technical phrase. No visitor asked that. It reads like a revision note from a regenerate-and-edit loop that got promoted into the page.

Check it: skim for duplicated headings and table rows, and read the FAQ for questions only the page’s author could have had.

One tell is noise. Three is slop.

None of the verdicts above rests on a single tell. Real people use templates, write slogans and forget alt text. What tips a page into slop is several tells stacking up with nothing pulling the other way. That is also why the highest AI score in the batch was 82, not 100: nearly all of these pages had real specifics, such as part numbers, prices and named sources, which each report credited under “What’s working.” Confidence stayed between 55% and 72% across the whole batch.

How those signals are weighed against each other, and why we publish a confidence level instead of a yes-or-no answer, is laid out in our public scoring methodology.

Check a site for slop in 30 seconds

The manual checks above take ten minutes per page. The free CrawlRanker AI slop detector runs them in about 30 seconds: it reads content, structure, imagery and tone, quotes the exact lines that triggered each finding, counts marker phrases, and gives a one-sentence fix for each. If it’s your own site that reads as slop, fix it in this order:

• Delete placeholders and duplicated blocks. That takes minutes.
• Replace generated images with real photos or product screenshots, and give every image alt text.
• Rewrite the headings as specific claims about your product, and cut the self-describing sentences.
• Re-scan, and compare the new report with the old one.

Keep going

• AI slop detector — free, no signup.
• How CrawlRanker scores a site — the full public scoring rubric behind every scan.

Or scan any URL with the free CrawlRanker SEO audit — seven SEO checks plus an AI-content score in about 30 seconds.