How do negative seed sites steer a lookalike search away?
Part of Lookalike websites
The short answer
A negative seed site is an example of a publisher the client does not want, given to a lookalike search so it moves away from that site's style or topic. Use one or two per unwanted trait, never more than the positive seeds, write down exactly what is unwanted, and test that known good sites still survive the search.
Key takeaways
- A negative seed describes a trait to move away from; a placement exclusion blocks a specific domain or page, and the two belong in separate fields of the brief.
- Negative seeds earn their place when positive seeds keep pulling in one recognizable kind of unwanted site.
- A standard information retrieval textbook rates positive feedback as much more valuable than negative feedback, so negative seeds should stay few and specific.
- A negative seed that shares useful subject matter with good sites can push those good sites out, so test with a control set.
- Each negative seed needs a written reason, an owner and a review date, or it becomes an unexplained rule nobody can challenge.
What is a negative seed site?
Definition
- Negative seed site
- A publisher, section or article given to a lookalike search as an example of what the client does not want, so the search ranks candidates lower the more they resemble it in the named trait.
Positive seeds tell a search what to look for. A negative seed tells it which neighbor to avoid. The idea comes from relevance feedback in information retrieval, where a searcher marks results as relevant or not and the system adjusts. The Rocchio method described in Manning, Raghavan and Schütze's Introduction to Information Retrieval (opens in a new tab) moves a query toward relevant documents and away from non relevant ones, with a separate weight for each direction.
A negative seed is not an exclusion. Google Ads lets advertisers exclude placements (opens in a new tab) "by domain, subdomain, path name, or individual pages," and an excluded placement does not serve. A negative seed blocks nothing on its own. It changes which candidates the search proposes, and a planner still reviews the results. Keep the two in separate fields of the brief: hard exclusions for places the client will never buy, negative seeds for traits the search should steer away from.
When should a lookalike search use negative examples?
Negative seeds solve one problem: positive seeds that keep attracting a recognizable kind of site the client rejects. If the first round of results is simply weak, better positive seeds are the fix, and the separate guide to seed site selection covers that choice. The table below sorts the situations a planner meets.
| Situation | Use a negative seed? | Better first step |
|---|---|---|
| Results share the right topic but in a style the client dislikes, such as gossip or clickbait | Yes, one example of that style | None; this is the core use |
| Results drift into a neighboring topic that shares vocabulary with the brief | Yes, one example of the neighboring topic | Also tighten the positive seeds to the exact subject |
| One specific domain keeps appearing and the client will never buy it | No | Add the domain to the hard exclusion list |
| Results are thin or off topic across the board | No | Replace or narrow the positive seeds |
| Results include unsuitable content such as violence or adult material | No | Set suitability rules for those categories |
| The client dislikes a site but cannot say why | Not yet | Ask what the site does wrong before using it |
How many negative examples help a lookalike search?
Fewer than the positive seeds. A standard textbook favors a light hand: Introduction to Information Retrieval notes that "Positive feedback also turns out to be much more valuable than negative feedback," and that some systems allow only positive feedback, which is the same as setting the negative weight to zero. A list of things the client dislikes can be longer and less coherent than a list of things it likes, and each extra negative example pushes the search in another direction.
Choose the narrowest example that shows the trait. If only a site's celebrity section is the problem, give a page from that section, not the whole domain. A whole site carries all the topics it covers, and the search will move away from all of them, including any the client wants.
Can negative examples remove good sites by mistake?
Yes, when the negative seed shares useful subject matter with the sites the client wants. Platform keyword exclusions show the same failure in a simpler form. Google's content suitability help (opens in a new tab) warns that "some words have multiple meanings" and that excluding a content keyword prevents ads from showing on content related to that word, "even in a different context than you intended." A negative seed carries far more words than a keyword, so the risk of pushing away a good site that discusses the same subject in a better way is real.
Feedback can also simply be wrong. A study by Kangasrääsiö, Chen, Głowacka and Kaski on concept drift and errors in relevance feedback (opens in a new tab) starts from the observation that a searcher's feedback can be noisy or mistaken and that intent shifts during a search; their system shows a timeline of past feedback so users can find and correct bad judgments. A negative seed chosen in a hurry is that kind of judgment, and it needs a way to be reviewed.
A control set test for each negative seed
- Name three known good sites. Pick sites the client approves that cover the brief's subject.
- Run the search without the negative seed. Note where each known good site ranks.
- Run it again with the negative seed. Note the new ranks and which candidates dropped out.
- Read the dropped candidates. Check whether they carried the unwanted trait or were good sites pushed out.
- Keep, narrow or remove the seed. If a known good site fell sharply, replace the negative seed with a narrower page.
How do you document negative seeds in a brief?
A negative seed without a reason becomes a rule nobody can challenge six months later. Gebru and coauthors' Datasheets for Datasets (opens in a new tab) proposes that each dataset come with a datasheet that "documents its motivation, composition, collection process, recommended uses, and so on." A negative seed list is a tiny dataset that shapes the results, and the same habits apply: why each example was chosen, what it shows and where it should not be reused.
| Field | What to write | Fictional entry |
|---|---|---|
| Example | The narrowest URL that shows the trait | The celebrity news section of a music site |
| Unwanted trait | Style, angle or neighboring topic, in one line | Gossip about musicians' private lives |
| Still wanted | What the example shares with good sites | Coverage of musicians and instruments |
| Category label | An IAB Tech Lab Content Taxonomy category if one fits | Pop Culture |
| Control result | What happened to the known good sites | All three kept their ranks |
| Owner and review date | Who asked for it and when it is re-read | Client brand lead; next brief |
The category label is optional but useful. The IAB Tech Lab Content Taxonomy (opens in a new tab) gives the industry a shared vocabulary for contextual targeting, so naming the unwanted topic in its terms lets another planner read the brief the same way. Style traits such as tone or ad clutter have no taxonomy category; describe them in plain words.
How does PlacementGraph handle what a client does not want?
What does a negative seed framework look like on a real brief?
Fictional example
A ukulele maker in Hawaii
An unnamed ukulele maker on Hawaii Island wants websites read by people learning and playing the instrument across Hawaii, and gives two positive seeds: a ukulele lessons site and an instrument craft magazine's luthier column.
The first round brings back two unwanted kinds of site. Some are music gossip pages that mention ukuleles when a celebrity plays one. Others are Hawaii vacation listicles that name ukuleles as souvenirs. A third group, lyric and chord aggregation pages, has thin text but no single unwanted subject.
The planner adds two negative seeds: one celebrity music news page for the gossip style and one souvenir listicle for the travel angle. The chord pages are left to the quality review instead. A control run confirms that a strumming tutorial blog, a luthier supply journal and a Hawaiian music history site keep their ranks. Both negative seeds go into the brief with their traits, the control result and the brand lead as owner.
Frequently asked questions
Should negative seeds be shown to the client?
Yes. The client should see each negative seed, the trait it stands for and the control result, because those choices shape which publishers never appear on the list. Showing them also catches mistakes: a client may recognize that an example was chosen for the wrong reason and ask for a narrower page.
Can a negative seed be a single article rather than a whole site?
It can, and a single article or section is the safer choice when only part of a site carries the unwanted trait. A whole domain carries all of its topics, so the search may move away from subjects the client wants. Pick the page that shows the trait clearly.
Do negative seeds carry over to the next campaign?
Only after a review. A negative seed reflects one brief's positive seeds and one client's dislikes at one time. Before reusing it, check that the trait is still unwanted, that the example page still shows it and that a fresh control run keeps the known good sites in place.
What if the client dislikes a site but cannot say why?
Hold it out of the search until the reason is clear. Ask the client to point at a page and describe what bothers them: the tone, the ads, the angle or the topic. If the answer is a specific domain they will never buy, it belongs on the exclusion list, not among the negative seeds.
Sources
- Introduction to Information Retrieval: The Rocchio (1971) algorithm (opens in a new tab)Christopher D. Manning, Prabhakar Raghavan and Hinrich Schütze, Cambridge University Press, 2008
- Interactive Modeling of Concept Drift and Errors in Relevance Feedback (opens in a new tab)Kangasrääsiö, Chen, Głowacka and Kaski, UMAP 2016 (arXiv)
- Exclude specific webpages and videos (opens in a new tab)Google Ads Help
- About content suitability (opens in a new tab)Google Ads Help
- Datasheets for Datasets (opens in a new tab)Gebru et al., Communications of the ACM (arXiv, revised December 1, 2021)
- Content Taxonomy (opens in a new tab)IAB Tech Lab
Researched and written by PlacementGraph Research. Read how we research, source and update these pages. Corrections: diana@placementgraph.com.