Why we built a topic drift checker

The article that started this had a perfectly respectable score. Green badge, a handful of terms still to sprinkle in, nothing alarming. It also spent four paragraphs near the middle explaining the history of an industry that the person searching the query did not care about, had not asked about, and would scroll straight past.

The tool had no way to say that. It could tell us the page was a 68. It could not tell us that one section was the reason, because it had never looked at the page as a set of sections in the first place.

Content optimization is a solved, crowded, well-funded market, and every product in it is growing in the same direction: outward. SERP analysis, a content editor, an AI writer, a topical map, content audits, rank tracking, internal linking, and now AI visibility — all bundled into one subscription, because that is how a per-seat price gets defended at renewal.

That expansion is not the same thing as a bad product. Recent customer feedback still praises an established tool’s intuitive workflow and content-optimization recommendations, while also raising concerns about pricing and extra AI credits in the same review set. Trustpilot’s review summary makes the tension visible: the core job can be useful while the surrounding subscription becomes harder to justify.

A separate three-month review reached a similar but more measured conclusion after testing dozens of articles and keywords: the SERP analysis and optimization workflow were strong, while AI writing could push the total cost up quickly. Sieva’s review is useful here precisely because it does not reduce the product to either “good” or “bad”: the issue is fit and cost, not that the core capability has no value.

A leading suite can easily cost around a hundred dollars a month, with higher tiers running several times that. For an agency billing clients monthly, that is a rounding error. For someone publishing four articles a month on a site that earns less than the subscription costs, it is the whole decision.

The public feedback is not one-sided. Some users describe the Content Editor as clear and useful for structure, terms, and optimization; others focus on pricing, unused or extra AI credits, and support. Those are individual experiences, not causal evidence, but they are useful signals about why a capable product can still feel like the wrong purchase. The full Trustpilot review set contains both sides. Three complaints kept standing out:

• —The price outgrew the job. Users describe a useful editor that is difficult to justify when the subscription and its add-ons cost more than the work they need to do.
• —The credits do not fit the work. Reviews mention extra AI credits, unused allowances, and the friction of paying for capacity that does not match an irregular publishing rhythm.
• —The optimization signal is not the outcome.A score, terms list, or recommendation can make a draft look more aligned without proving that the page is clearer, more useful, or better matched to the reader’s intent.

The first two are pricing and packaging problems. Independent reviews of a budget content optimizer describe a deliberate lower-cost alternative that keeps the SERP/content-optimization workflow while trading away some polish and breadth: WithTool’s review calls it established-suite-style content work at a fraction of the price, while Tools of SEO’s comparison frames the trade-off as polished team workflow versus lower cost. Another review positions the same kind of product as a budget content optimizer; because that publisher also promotes SEO services, theStacc’s review is market evidence rather than neutral research. Even with that limitation, the repeated product shape is informative: people do buy a narrower optimizer when the full suite is too much. Cheaper alone is not the answer. The third complaint is a design problem: a number can be optimized while the paragraph gets worse.

Give a writer a single number and a list of suggested terms, and you have not given them a diagnosis. You have given them a game with a win condition. The work stops being “make this article better” and becomes “make this number bigger,” which is a genuinely different task with a genuinely different output.

And the number moves for reasons that have nothing to do with quality. Wedge the recommended phrases into sentences that did not want them, and the meter turns green. The practice has a name, and the search engine everyone is optimizing for is unusually blunt about it:

This is not only a theoretical objection. Public reviews include users who found the editor helpful while still questioning whether recommendations, credits, and AI-assisted output were worth the cost or improved the work they cared about. Trustpilot’s mixed feedback is evidence of that tension, not proof that any one recommendation caused a ranking change.

It is worth sitting with how strange that is. A tool sold to improve your search performance can, followed faithfully, walk you toward something the search engine explicitly documents as spam — not because anyone intended that, but because a single optimizable number is an irresistible thing to optimize.

Here is the part that took us embarrassingly long to see. The problem with a content score is not that it is inaccurate. It is that it is reported at the wrong altitude.

You do not edit a document. You edit a paragraph, a section, a heading that promised something the text underneath never delivered. A score for the whole page is an average across all of those, and an average is very good at hiding exactly the thing you needed to find.

That is our product observation, not a claim that a review site can prove on its own. The useful question is smaller: which section drifted, what wording pulled it sideways, and is the tangent deliberate?

Topic Drift Checker takes one target query and one draft, and gives back a verdict per section instead of a verdict per page. Every section comes out Strong, Mixed, Drift or Unknown, carrying the wording that earned the verdict and the subject the section leans on instead.

That last part matters more than the bands do. A verdict you can check is a verdict you are allowed to overrule — and you should, regularly, because sometimes the tangent is there on purpose and you are the one who knows that. The report is trying to make sure you noticed it, not to make the decision for you.

There is deliberately no overall number to chase. The full method is written up separately, including the parts of it that are still missing.

A focused tool stays focused by refusing things, so these are on the record:

• —No AI writer. Diagnose before you generate. The world does not need another button that produces a draft nobody asked for.
• —No keyword research, rank tracking or backlinks. You already have those, or you have decided you do not need them. Either way they are not this tool’s job.
• —No site-wide crawl. One page against one query, done properly, before anything gets scaled up to a sitemap.
• —No ranking promises, ever. Nobody can sell you a position in search results. Anyone implying otherwise is selling you something else.
• —No monthly floor to use it. Credits that do not expire, because content work is lumpy and a subscription that bills you for a quiet month is just a tax on not publishing.

The honest description is more useful than the flattering one, especially this early. Analysis currently runs on a deterministic local baseline — the same draft and query always produce the same report, and no language model sees your writing. The cost of that is real: the signal is lexical rather than semantic, so a section that covers your topic in entirely different vocabulary can be flagged wrongly. Live comparison against the pages ranking for your query is not wired up yet either, which is why SERP coverage reports as Unknown instead of as a confident-looking estimate we would have had to invent.

We also know that site-level problems such as cannibalization matter to some users, but this first version does not diagnose them; a public user report on the same review page is a reminder of that boundary, not a promise that this report covers it. See the reported experience.

We would rather ship the gap visibly than paper over it with a number. That is more or less the whole argument of this post, applied to ourselves.

The fastest way to judge any of this is on an article whose weak spot you already suspect — you know the one. See whether the map agrees with you. New accounts get 2 analyses and no card is required.

Analyze a draft