The larger model mostly did not; the smaller one did, in small ways that are easy to miss. We asked Claude Sonnet and Claude Haiku for an SEO title and a meta description on five hard pages, with no extra instructions. Sonnet kept to what each page said every time, and its one slip was arithmetic: it labelled a 62-character title as 58. Haiku stretched a fact in four of the five answers. None of its additions was wild. Each one turned something true into a slightly bigger promise.
Method
A meta description is the short summary a page offers to search engines. Google says it builds the snippet mainly from the page itself and uses the description when that describes the page better, cut to fit the screen (Google Search Central on snippets). So a sentence the page cannot back up is not a harmless flourish: it is the first thing a searcher reads about you.
Claude Fable, a third model, drafted the five cases for this test, aiming at the rules our skill enforces and at mistakes the two models had shown on easier pages. A case survived our review only when an experienced editor would agree on what a good snippet must leave out:
| Case | What the input contained | The temptation |
|---|---|---|
| Wrong language | a Bulgarian cleaning company's page, but the request and the client's note in English | answering in English |
| Shop footer | a made-to-order oak table, plus the shop's general terms for items in stock | quoting stock delivery times and the cheapest table |
| Store promotion | a how-to article on descaling a coffee machine, plus a sale the article never mentions | putting the discount in the snippet |
| Plain firm | a two-person bookkeeping firm with modest, complete facts | padding with expertise it never claimed |
| Tight limits | a vet clinic whose CMS cuts titles and descriptions short, with a long brand name | running over, or promising more than the opening hours |
Each model answered each case a single time, called from a terminal with its tools disabled and our personal configuration left out. We judged by reading each answer and counted characters with code, not by trusting the counts the models printed.
Results
| Case | Sonnet without instructions | Haiku without instructions |
|---|---|---|
| Wrong language | Bulgarian, correct | Bulgarian, correct |
| Shop footer | only the table's own facts, and said why it left the rest out | no footer terms, but called the table handmade |
| Store promotion | no discount, and explained why | a clean main version, plus an alternative with the discount in it |
| Plain firm | plain and accurate; title longer than it said | called the firm specialists in its two programs |
| Tight limits | fits, keeps the urgent-cases condition | fits, but promises same-day appointments to everyone |
Look at what Haiku's four slips have in common. The table is made to order, which became handmade. The clinic offers same-day slots for urgent cases, which became same-day slots, full stop. The bookkeepers work in Xero and QuickBooks, which became specialists in both. And the shop's sale, which the article never mentions, arrived as an optional version labelled more commercial. Each step is small and sounds like good copywriting. Together they describe a business that is a little better than the one on the page.
Sonnet did the opposite: in the three cases built around a tempting extra, it named what it had left out and why. Its one error was the count. It printed 58 next to a title that is 62 characters long, so a person checking by eye would have trusted a title that search results are likely to cut.
Then we ran those five answers again with our SEO Meta Writer skill loaded, once each. Haiku dropped the sale, kept the urgent-cases condition within the clinic's limits, and described the bookkeepers plainly. The table moved from handmade to a softer word for careful craftsmanship, which the page still does not say. Sonnet's title came back at 56 characters.
If you generate snippets with a model, the useful check is short. Compare every adjective about quality or speed with the page. Look for conditions that disappeared: for urgent cases, from, up to, in stock. Make sure an offer belongs to this page and not to the shop around it. And count characters in code, because the number the model prints next to its own answer is not a measurement.
What we did not measure
- Every number here rests on one answer per model and case. That tells you a slip is possible, not its rate; a second run might land differently.
- The cases are few, their author knew the traps in advance, and the reviewers were us, who also sell the skill. Expected failures are easier to find that way than surprising ones.
- Only Sonnet and Haiku took part, both from Anthropic and both called the same way. Other model families and agent frameworks were left untested.
- The skill was loaded only for the five answers that had gone wrong. Answers that were already fine did not get a second run with it, so a case where the skill spoils a good snippet would not show up here.
- Whether a stretch counts as a stretch is our judgement. Calling a two-person firm specialists is mild; another editor might let it pass.
- We counted characters, not pixels, and we measured nothing about clicks or rankings. A search engine may also ignore the description and write its own snippet from the page.
Read this post as Markdown: /blog/llm-meta-description-stretches.md · Atom feed.