How the drift check works.

One target query and one draft. No project setup, no crawl, no keyword list to import. Topic Drift Checker assumes you already chose the query — from Search Console, a rank tracker, or a client brief — and that what you need now is a second opinion on the draft.

You can paste the text or give it a published URL. Importing a URL is free and never costs a credit, so a failed fetch costs you nothing.

Sections are split at your headings. Before anything is scored, the parser drops the furniture that is not really your article — breadcrumb preambles and trailing Related posts or Keep reading blocks — so recommendation widgets cannot drag a verdict down.

Each section is compared with the target query and given an alignment figure, which falls into one of four bands:

• —Strong— 75 and above. The section’s wording lines up with the query.
• —Mixed — 50 to 74. Related, but the emphasis has moved. Worth a read.
• —Drift — below 50. Little direct query wording, and the dominant terms belong somewhere else.
• —Unknown — under about 50 words, or a heading that only means something in context (“Short answer”, “This one”). There is not enough to judge, so no verdict is invented.

Every flagged section carries the excerpt that produced the flag and the terms the section is actually dominated by. The claim is always checkable against your own text.

A single number turns editing into a game with a win condition. The well-documented result is writers stuffing suggested terms into sentences that did not want them, until the meter turns green and the prose reads worse than the draft they started from.

A report here shows the section, the verdict, and the evidence. Where a dimension cannot be supported, it says Unknown rather than filling the space with a plausible number. That is the whole design.

This is an early product and the honest description matters more than the flattering one.

• —No live search-results comparison. Nothing is currently measured against the pages ranking for your query, so SERP coverage reports as Unknown on every report.
• —The signal is lexical, not semantic. Analysis matches wording, not meaning. A section that covers your topic in entirely different vocabulary can be flagged — which is why every verdict ships with its evidence for you to overrule.
• —Overall alignment, intent match and section consistency are not scored. They appear as Unknown until there is evidence good enough to stand behind.
• —English only, one query at a time. Analysis assumes a US English context.
• —No rerun comparison. You can analyse an edited draft, but the tool will not diff it against the previous report for you yet.

The upside of that narrowness is real: no language model sees your draft, and the same draft and query always produce the same report.

The fastest way to judge any of this is on an article whose weak spot you already suspect. New accounts get 2 analyses and no card is required.

Analyze a draft