Verify AI content against a source document

Define what the reader should be able to do

Replace a broad topic such as “marketing reports” with a specific situation: a manager receives a spreadsheet every Monday but cannot tell which decision it should change. The intended outcome is that the reader can choose a few useful metrics, connect them to actions, and remove rows that contribute nothing. That outcome provides a structure before anyone begins drafting.

Record the reader’s role, existing knowledge, and next action. Include boundaries: this article will not teach advertising API integration or compare every business intelligence platform. Those exclusions protect space for the actual explanation. A focused article can finish one useful conversation without trying to become a complete course on marketing operations.

Research intent without copying the search results

Review results for the main question and two related formulations. Note the types of pages that appear, the tasks they cover, and the steps they leave unexplained. Look for missing calculations, undefined terms, or examples that combine incompatible situations. Repeated competitor headings are research inputs, not an outline you must reproduce.

Document the research date, language, and target market. Leave keyword volume unmeasured unless a tool actually supplied it. Search suggestions and the number of results do not establish demand. Check your own content inventory as well: a guide to selecting weekly metrics should have a different purpose from an existing tutorial on building a dashboard.

Build an evidence map before generating prose

Create a row for each important claim with its source, verification date, supported conclusion, and limitation. Product documentation establishes a feature or requirement. A company export establishes an observation for that company and period. Those are different forms of evidence. Mark an invented teaching example as illustrative and keep its inputs available for review.

Google’s content guidance emphasizes usefulness and reliable evidence rather than a preferred word count. Translate that into an editorial rule: missing information remains a question for the expert. Do not ask a model to add the author’s firsthand experience when the source packet contains none. A clearly explained method and a labeled example can provide value without a fabricated client story.

Worked example: a coffee brief with an unanswered question

This is an original teaching example inspired by the Practicum lesson “Only from this document.” Sáng Coffee is a fictional training brand, not a client or a verified business result. The source packet below was written for this demonstration. It contains voice and photography guidance; it deliberately contains no partner-logo policy.

Source packet, version 1: “Write in plain, welcoming language. Describe the drink and its preparation without health claims. Use close photographs of the drink in natural light. The brand owner must approve product information before publication.” Those four sentences are the complete evidence available here.

Assignment: write a caption about the iced coffee and answer whether a reseller may print the logo on its own packaging. A deliberately flawed demonstration answer says: “Our iced coffee boosts focus. Partners can use the logo if they leave clear space.” Neither sentence follows from the packet. Fluency and plausible brand conventions do not supply the missing permission or health evidence.

Use this instruction: Answer only from the source packet. For each factual claim, quote the supporting sentence. If the source does not answer, write “Not in the document” and name the missing input. Check the quotations yourself: an instruction reduces ambiguity but does not guarantee that a model follows it.

Accepted revision: “An iced coffee, photographed in natural light. Product details require owner approval before publication.” Partner packaging: “Not in the document. Ask the brand owner for an approved partner-logo policy.” This is a grounded draft, not approval to publish. No model run or measured error-reduction result is claimed.

To repeat the exercise, replace the four source sentences with your own approved brief, keep one answerable and one unanswered question, and fill the worksheet. The useful result is a claim-to-evidence map and a specific request for missing information.

Download the complete fictional source packet · Download the completed source-check worksheet · Try the guest practice

Make acceptance criteria concrete

Suppose the article uses an illustrative funnel with 120 inquiries, 30 sales-accepted leads, and six closed deals. The acceptance rate is 25%; the close rate among accepted leads is 20%. A writer must state the denominators. “Conversion improved” is not an acceptable substitute when the reader cannot tell which stage is being measured.

The brief below ties an editorial request to an observable result. Add your source links or a “source required” note to the working document. Use aggregate examples when individual customer details are unnecessary. If the article later discusses profitability, specify the period, currency, and included costs before asking the writer to draw conclusions.

Separate facts from generation instructions

Put verified inputs in one block and editorial preferences in another. The first contains evidence; the second contains tone, structure, and formatting. Ask for a missing-information list before an outline, and review the outline before drafting. When an unsupported number appears, return it to the questions list rather than accepting it because it makes a paragraph sound convincing.

The site’s brief-generator helps organize project inputs and flag unanswered questions. For an article assignment, add the evidence map and acceptance criteria described here. The tool does not replace search research or authenticate every source. Treat its output as a document draft that an editor must adapt to the specific reader and decision.

Review the deliverable rather than a detector score

Read the draft backward from its promised outcome. Could a reader complete the task without searching elsewhere for an omitted step? Trace numerical conclusions to their inputs and product instructions to current documentation. Then read the introduction alone: it should answer the question and establish relevant limits instead of explaining why the subject is important.

Common failures include treating a keyword list as a brief, requiring the same outline for every intent, and demanding unsupported personal anecdotes. Another is using an AI detector score as the acceptance standard. No reliable promise of detector evasion replaces editorial review. Archive the evidence packet, verification date, and adjacent topics intentionally excluded from the finished article.