Many queries are not new investigations but attempts to relocate something the user has already seen. The system distinguishes re-finding from new-finding intent and routes the ranking accordingly, surfacing prior results for re-finders and fresh results for new-finders.
Patent Overview
- Inventor
- Jaime Teevan, others
- Assignee
- Microsoft Corporation
- Filed
- 2008-04-25
- Granted
- June 17, 2014
The Challenge
The Challenge
A large fraction of queries are repeats. The user is not investigating something new; they are trying to find again a page they already visited. The challenge: detect re-finding intent and route the ranking to favor the previously visited result, without breaking the new-finding case where the user wants fresh information on a familiar topic.
- The Same Query String Has Two Intents — Per query, the same string can mean 'show me what I saw before' or 'show me something new on this topic'.
- Result Volatility Punishes Re-Finders — Per ranker cycle, fresh content rotates into the SERP and pushes the previously visited result down, breaking re-finding without serving new-finding any better.
- Personalization Alone Does Not Solve It — Per user, a personalized ranker that knows the user's interests still cannot tell whether this query is a return visit or a new investigation.
- Bookmarks Capture A Fraction Of Returns — Per user, explicit bookmarking covers a small share of actual return visits. Most re-finding happens via re-issued queries.
- Mis-Routing Frustrates In Both Directions — Per query, surfacing old results to a new-finder feels stale; surfacing new results to a re-finder feels broken.
Innovation
How The System Works
The system observes the user's query and click history, classifies a new query as either re-finding or new-finding intent based on prior interaction signals, and routes the ranking to favor previously visited results for re-finders or fresh content for new-finders. Re-finding can include rank-stabilization so the previously clicked result holds its position across visits.
- Detect Query Repetition — Per user, the system checks whether the query, or a close variant, was issued previously and produced a click.
- Score Re-Finding Probability — Per query, features including prior query match, prior click match, time since last visit, and query specificity feed a classifier.
- Identify Candidate Re-Finding Targets — Per user, previously clicked results for this query or related queries are flagged as re-finding candidates.
- Apply Rank Stabilization — Per re-finding candidate, the result is boosted so it appears at or near its previous SERP position.
- Preserve New-Finding Slots — Per query, a portion of the SERP remains available for fresh content so users who actually want something new still see it.
- Track Re-Finding Outcome — Per query, whether the user clicks the previously visited result or skips to fresh content updates the re-finding classifier.
- Adapt Mix Over Time — Per user, the balance between re-finding boost and new-finding freshness adapts based on observed behavior patterns.
Repeat Queries Are A Distinct Intent Class
The patent's load-bearing idea is that repeat queries are not a noisy version of new queries. They are a separate intent class that requires different ranking behavior. Treating repeats as new investigations breaks the user's actual goal.
Intent-Conditional Ranking For Repeats
Per query, the ranker reads whether the user is re-finding or new-finding and routes the ranking accordingly. Per user, this judgment is informed by personal query and click history.
- Repetition Detection — Per query, the system identifies prior issuance and prior clicks.
- Re-Finding Classifier — Per query, features predict whether this is re-finding or new-finding.
- Stabilization And Freshness — Per SERP, re-finding boost is balanced against fresh-content slots.
Technical Foundation
Technical Foundation
The patent specifies query repetition detection, re-finding probability scoring, rank stabilization for prior results, and balance with new-finding freshness.
- Per-User Query History — Per user, prior queries and clicks are stored as the foundation for re-finding detection.
- Repetition Signal Set — Per query, signals include exact match, near match, paraphrase, and topical overlap with prior queries.
- Re-Finding Classifier Features — Per query, classifier features include query repetition signal, prior click strength, time since last visit, query specificity, and ambiguity.
- Rank Stabilization Mechanism — Per previously clicked result, a boost preserves the result's position in the current SERP when re-finding intent is detected.
- Freshness Slot Reservation — Per SERP, a portion of slots stays available for fresh content so the system handles new-finding even when re-finding probability is high.
- Outcome Feedback Loop — Per query, the user's click behavior on this SERP updates the classifier for future re-finding judgments.
The Process
The Process
From a query, the system checks the user's history for prior issuance and clicks, scores re-finding probability, applies rank stabilization to prior results when probability is high, and preserves freshness slots for new-finding cases.
- Receive Query And User Identity — Per query, the query string arrives with a signed-in user identifier or session identifier.
- Look Up Prior Query History — Per user, prior queries and clicks are retrieved from the user's history store.
- Score Re-Finding Probability — Per query, classifier features produce a probability that the user is re-finding.
- Flag Re-Finding Candidates — Per user, previously clicked results for this query or close variants are identified.
- Apply Rank Stabilization — Per re-finding candidate, the result receives a stabilization boost based on re-finding probability.
- Reserve Freshness Slots — Per SERP, slots are kept available for fresh results to handle new-finding cases.
- Update Classifier From Outcome — Per query, the user's click behavior updates the re-finding classifier for future queries.
Quality Control
Quality Control
Re-finding boost can stale the SERP, miss genuine new-finding cases, or stabilize a result that is no longer the user's actual target. The patent specifies safeguards to keep results responsive.
- Probability Threshold — Per query, rank stabilization is applied only when re-finding probability exceeds a confidence threshold.
- Time-Decay Of Prior Results — Per prior click, stabilization weight decays with time since the click so very old visits do not dominate current results.
- Mandatory Freshness Slots — Per SERP, fresh-content slots are reserved regardless of re-finding probability so new content stays discoverable.
- New-Finding Signal Override — Per query, explicit new-finding signals such as query modifiers like 'new' or 'latest' override the re-finding boost.
- Cross-User Validation — Per result, stabilization boost is sanity-checked against global signals so a result that has materially decayed in quality is not preserved at the user's expense.
Real-World Application
Re-finding versus new-finding routing powers personalized SERP stability across return visits. A user who searched 'best chrome extensions for designers' last week and clicked the third result sees that same page surfaced when they issue the same query this week, even though the global ranking has changed. A user who searched the same query for the first time sees the fresh global ranking.
- Distinct intent Classification — Re-finding and new-finding are treated as separate query intent classes.
- Per-user history Signal Source — Prior queries and clicks drive re-finding detection.
- Balanced SERP Result Mix — Stabilized prior results combine with fresh-content slots.
Why Repeat Queries Are A Large Share Of Search
Per user, a substantial fraction of all queries are repeats of prior queries. The system that ignores this and reranks fresh on every visit treats every user like a first-time investigator and breaks the return visit case.
Why Brand Queries Often Stabilize First
Per query type, brand queries from prior visitors are almost always re-finding. The system treats them with strong stabilization, which is why navigational SERPs feel almost static across visits while informational SERPs rotate more.
<\/section>What This Means for SEO
What This Means for SEO
Re-finding routing means that once a user has visited your page from a query, the engine actively works to surface your page again for that user on the same query. The strategy implication is that earning the first click is more valuable than the per-query position suggests, and breaking the re-visit experience costs the re-finding signal.
- First Clicks Compound Across Future Queries — When a user clicks your page from a query, the engine increases the probability of surfacing your page again to that user on the same query. Earning the first click is not a single transaction; it is purchasing future ranking stability for that user.
- URL Stability Protects The Re-Finding Signal — If the URL the user clicked breaks or 404s, the re-finding boost cannot route the user back. Stable canonical URLs preserve the per-user ranking stability that has already been earned.
- Title And Meta Stability Aid Recognition — Re-finders recognize the result by title and snippet as much as by URL. Changing the title or meta description on a high-traffic page can mute the recognition signal even when the URL is stable, slowing re-finders' click decision.
- Brand Queries Carry Re-Finding Weight — Brand and product navigational queries are dominated by re-finding intent. Owning the canonical brand page with consistent metadata across years is a re-finding moat, not just a brand SEO basic.
- Returning Visitors Are A Ranking Asset — A page that pulls return visitors accumulates the personalized stabilization signal across its audience. Investing in content that is worth returning to compounds in the per-user ranking, not just in direct return traffic.
- New-Finding Slots Reward Freshness — The freshness slots reserved on every SERP go to fresh content. Pages that publish meaningful updates earn the new-finding traffic alongside the stabilized re-finding traffic the existing leaders hold.
- Personal SERPs Are Personal Histories — The SERP a return visitor sees encodes their prior interactions with your domain. Disrupting that history with redirects, URL migrations, or content removals does not just hurt rankings; it severs the re-finding signal the engine has spent visits learning.