Zhihe GrowthAI citations may appear as small passage-end markers, source cards, image or product blocks, or a Sources list at the bottom. Some platforms reveal destination URLs only on click or hover; others route through an intermediary URL. Copying only visible text may lead reviewers to treat a title as a complete source or wrongly record no URL. That undermines subsequent citation-rate and quality audits.
Keep Four Types of Evidence for Each Question
- The complete question and answer boundaries, including web access, deep-thinking or expert mode, platform, date, and region.
- The full native answer, without screenshots cutting off its latter half. Cross-check text records against screenshots.
- Each visible source's title, location, type, and target URL revealed by expansion or hover.
- The final destination URL after clicking, the redirect chain, and the passage that actually supports the claim.
If a source is an internal product card or file without a public webpage URL, record its type and access method rather than guessing an address. Preserve original links even when they return 404 or require login, and mark the access status. One answer may cite several sources; a brand may also appear in a summary without any official-site citation. Keep these cases distinct.
Common Interface Traps and How to Handle Them
| What You See | Common Misclassification | Correct Verification |
|---|---|---|
| Sources +2 or a small citation chip | Only one source | Expand the list and open each source |
| Visible Page Title | URL recorded | Hover or click to retrieve the actual address |
| Platform redirect link | The platform is the original source | Continue to the final landing page |
| A citation midway through a long answer | The whole paragraph is supported by the same page | Compare the associated sentence with the original passage |
| Recommended product card | An organic official-site citation | Record the card category and actual destination |
Each screenshot should cover one complete question and complete answer. For long pages, preserve a native full-length screenshot and split it at verified Q&A boundaries, but do not combine adjacent questions into one result image. Sidebars and unrelated controls are not answer evidence. Text records preserve URLs and classification reasons; screenshots preserve the platform's appearance at that moment. Both matter. For lightweight retests that do not require screenshots, still retain the complete answer and final URLs.
How Zhihe Growth Uses Recovered Source URLs
Zhihe Growth's measurement method separates mentions, visible citations, and factual accuracy. If a record says no URL despite a visible source title, return to the original conversation and expand its sources rather than infer absence from a screenshot. Public case reports use only verified, scoped interim results. The SuperPDR and ELEREIN named case studies do not disclose clients' complete private question sets.
After recovering URLs, perform a source-support audit: does the link point to the official site, does the page support the answer, and were any conditions omitted? Bing's official AI Performance documentation describes aggregate citations and topic signals, not all native conversations. Dashboard charts cannot reconstruct the exact sources for a particular exchange. Keep not collected, no visible source, and source inaccessible as separate statuses to avoid overstating or understating citations.