What is lookalike drift, and when should expansion stop?
Part of Lookalike websites
The short answer
Lookalike drift is the gradual slide of a site search away from its seed sites' theme as each expansion round uses the previous round's finds as new examples. Planners detect it by scoring only new sites against the original seeds, stop after two rounds with no new on theme sites, and reset by returning to the last clean round.
Key takeaways
- Drift starts when a round's results, not the original seeds, become the examples for the next round.
- Bridge topics that share vocabulary with the brief carry a search into neighboring subjects one step at a time.
- Score each round on new sites only, against a written description of the original seeds, or repeats will hide the slide.
- A bounded stop rule, such as two rounds with no new on theme sites, keeps expansion from running on habit.
- Resetting means going back to the last clean round and the original seeds, not patching the latest list.
What is lookalike drift?
Definition
- Lookalike drift
- The movement of a lookalike site search away from the theme of its original seed sites over successive expansion rounds, so later rounds return sites that resemble earlier results more than they resemble the seeds.
Information retrieval researchers call the same effect query drift. Manning, Raghavan and Schütze's Introduction to Information Retrieval (opens in a new tab) explains it with pseudo relevance feedback, where a system assumes its top results are relevant and expands the query from them: "if the query is about copper mines and the top several documents are all about mines in Chile, then there may be query drift in the direction of documents on Chile." Nothing in the first round was wrong. The second round simply learned the wrong lesson from it.
Lookalike site searches repeat that pattern when a planner, or a tool, takes the best sites from round one and uses them as seeds for round two. Each round inherits the quirks of the last. A site that matched the seeds for a secondary reason becomes an example, and the next round matches that secondary reason more strongly.
Why do lookalike expansion rounds wander off theme?
Drift is a property of the process, not of any single bad site. The table names four mechanisms a planner can look for in the round history. They are separate from the pitfalls covered on the lookalike websites hub, which describe problems inside a single result set.
| Mechanism | What happens between rounds | Where it shows in the record |
|---|---|---|
| Chained seeds | Round two uses round one's results as examples instead of the original seeds | The seed list grows each round while the original seeds carry less weight |
| Bridge topics | A neighboring subject shares words or readers with the brief and pulls the search across | New sites cover the neighbor first and the brief's subject second |
| Secondary trait amplified | A site matched for its tone, region or format, and that trait becomes the target | Later sites share the trait but not the subject |
| Thin sample | A site joined on one or two matching articles and then served as an example | A seed whose sampled articles mostly cover something else |
How can a planner detect lookalike drift?
Measure each round against a fixed reference, and count only what the round added. The second part matters because repeats flatter a round. In evaluating relevance feedback, Introduction to Information Retrieval describes using the residual collection (opens in a new tab), "the set of documents minus those assessed relevant," for the second round of evaluation. Applied to site lists, that means scoring only the sites a round found for the first time.
Signs of drift to check after each round
- Fewer than half of the round's new sites match the written description of the original seeds.
- The round's new sites match the previous round's sites better than they match the original seeds.
- A subject that was absent from the seeds now appears in several new sites' sampled articles.
- The IAB Tech Lab Content Taxonomy categories of new sites differ from the categories recorded for the seeds.
- New sites raise suitability flags that earlier rounds did not.
- The planner can no longer explain a new site's fit by pointing to one of the original seeds.
The taxonomy check is cheap and catches large moves. The IAB Tech Lab Content Taxonomy (opens in a new tab) gives a shared vocabulary for contextual targeting, so recording the seeds' categories at the start gives a fixed rubric for later rounds. It misses small moves inside one category; the written description of the seeds catches those. The half match threshold in the checklist is this page's proposal, not a published standard; set the number in the brief and keep it fixed.
When should lookalike expansion stop?
Expansion needs a stop rule written before the first round, or it runs until the list looks big enough. Research on interactive search gives a concrete model. In a TREC 2016 dynamic search task, Di Buccio and Melucci of the University of Padua tested two stopping criteria (opens in a new tab): a fixed number of feedback iterations, or stopping "after 2 consecutive iterations of feedback with no relevant documents in the top 5 returned results" or at the iteration cap. That was an experimental setting, not an industry rule, but its shape transfers well.
Stopping is also right when the list is big enough for the budget, even if rounds still find good sites. A plan that needs forty sites does not improve because a fifth round found a forty first. Record why expansion stopped, so the next planner knows whether the search was exhausted or simply complete.
How do you reset a drifting lookalike search?
Do not patch the latest list by deleting the obvious misfits. The sites that caused the drift can be plausible finds from earlier rounds, and if they stay in place they can cause it again. Research on relevance feedback supports looking back: Kangasrääsiö, Chen, Głowacka and Kaski's work on concept drift and errors in relevance feedback (opens in a new tab) adds a timeline of past feedback so a searcher can find and correct the judgments that led the system astray.
Five steps to reset a drifting search
- Find the last clean round. Read the round history and mark the last round whose new sites still matched the original seeds.
- Identify the bridge sites. List the sites from that round onward that carried the neighboring subject into the search.
- Return to the original seeds. Restart from the seeds and the clean rounds' results, without the bridge sites as examples.
- Tighten the description. Add one line to the seed description naming the neighboring subject that pulled the search away.
- Rerun with the stop rule. Score new sites only, and stop at two empty rounds or the cap.
Some drifted sites may still be worth testing as adjacent ideas. Keep them apart from the core list with the reason they are there, as the guide to adjacent audiences describes, rather than letting them back in as seeds.
How does PlacementGraph compare candidates with the client's examples?
What does lookalike drift look like on a real brief?
Fictional example
An indoor herb garden kit maker in Iowa and Nebraska
An unnamed Des Moines maker of countertop hydroponic herb garden kits wants websites read by home cooks in Iowa and Nebraska. The seeds are a kitchen herb growing blog and a home cooking site's fresh herbs column. The seed description reads: articles about growing and using culinary herbs indoors at home.
Round one adds kitchen gardening and herb care sites, all on theme. The planner then uses round one's best finds as new seeds. Round two adds hydroponics sites in general, and round three adds commercial greenhouse trade press and two home grow sites that the client's suitability settings exclude. Scored on new sites only, round three matches the seed description for one site in six.
The planner resets to round one, notes that general hydroponics was the bridge topic and adds a line to the description: home kitchen use, not crop production. The rerun adds cooking and herb sites for two rounds, then two empty rounds end it. The general hydroponics sites go to an adjacent ideas tab with their reason.
Frequently asked questions
Is lookalike drift always a problem?
No. Drift is a problem when it is invisible. A planner who sees the search moving toward a neighboring subject can decide to test that subject on purpose, with its own budget line and reason. The harm comes when drifted sites enter the core list labeled as matches to the client's favorites.
Does drift happen in category based searches too?
Yes, in a different form. A category search does not chain seeds, but a planner who widens from one category to its parent and then to sibling categories moves the list in steps. Recording the original categories and scoring each widening against them works the same way as scoring rounds.
Can a campaign drift after the list is approved?
The list itself does not change, but delivery can widen. Google Ads notes that when placements are combined with topics or keywords, ads can show on content matching any of them, which reaches a broader set of pages than the approved sites. Check the targeting settings, not only the list.
Should the round history be shared with the client?
A short version helps. Show the original seeds, the number of rounds, the stop reason and any reset with its bridge topic. A client who sees why a search stopped has a reason not to ask for one more round that would bring back the sites the reset removed.
What should a round log record for each expansion round?
Record the examples used, the date, the sites found for the first time, how many of those matched the original seed description, any new taxonomy categories or suitability flags, and the decision taken: continue, stop or reset. One line per round is enough to show later where a search started to slide.
Sources
- Introduction to Information Retrieval: Pseudo relevance feedback (opens in a new tab)Christopher D. Manning, Prabhakar Raghavan and Hinrich Schütze, Cambridge University Press, 2008
- Introduction to Information Retrieval: Evaluation of relevance feedback strategies (opens in a new tab)Christopher D. Manning, Prabhakar Raghavan and Hinrich Schütze, Cambridge University Press, 2008
- Evaluation of a Feedback Algorithm inspired by Quantum Detection for Dynamic Search Tasks (opens in a new tab)Emanuele Di Buccio and Massimo Melucci, University of Padua, TREC 2016 (NIST)
- Interactive Modeling of Concept Drift and Errors in Relevance Feedback (opens in a new tab)Kangasrääsiö, Chen, Głowacka and Kaski, UMAP 2016 (arXiv)
- Content Taxonomy (opens in a new tab)IAB Tech Lab
- Placement targeting: Definition (opens in a new tab)Google Ads Help
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.