Documentation / The Data / AI Model Answers
AI Model Answers
See which sites AI models recommend in your market, which searches they suggest, and which phrases they associate with its content.
What is collected
When a campaign is finalized, OppAlerts asks several AI models three types of questions about the campaign's industry:
- Which sites they would recommend in this market. This is Competitor LLM Recs.
- What people search for in this market. This is the LLM Search Keywords tab of Keyword Research.
- Which phrases the market's pages use. This is the LLM Content Phrases tab of Keyword Research.
The models answer without web search, using what they learned during training. This is not a search ranking, and it is not what an AI tool cites today.
How the answers are scored
- Each appearance contributes the model's weight divided by the item's position in that answer raised to the power 1.5.
- An item named first contributes much more than one named tenth. At position 4, the model's weight is divided by 8.
- Score is the sum of those contributions. Seen is the raw count of times named. A site named tenth in every answer has a high Seen and a low Score: the models know it but do not lead with it.
- Models counts how many different models named the item. The response does not identify the models.
- The score is calculated when you read the data. It is never stored, so a change to a model's weight affects it immediately.
How the answers are merged
- Sites are merged by host name, in lowercase and without
www. One site is one row. - Keywords and phrases are merged by normalized text: lowercase, no punctuation, and single spaces.
formslists each written form that was merged into the item, with its count.
One item
A full sample item returned by the llm_answers_ tools:
{
"item": "example-hotels.com",
"agreement_score": 4.2131,
"appearances": 9,
"models": 6,
"best_rank": 1,
"forms": {
"example-hotels.com": 5,
"Example Hotels": 4
}
}
item: the site, keyword, or phrase.agreement_score: the position score.appearances: the Seen count.models: how many different models named it.best_rank: its best position in any answer.forms: each written form, with its count.
How to read your position
- Absent from every model: the models do not connect you with this market. Backlinks and third-party coverage change that.
- Named by one model only: that model often learned it from one strong source.
- Named by all, always low: the models associate you with the market but do not make you their default answer. Look at what the sites above you have.
- A top site that is not a competitor is usually a publisher, a directory, or a review site. Those are the pages to be on. Take them to Opportunity Types.
- Sites from the wrong market: the campaign's industry is wrong. The question names the industry.
In MCP and the API
llm_answers_top_sites,llm_answers_search_keywords, andllm_answers_content_phrases.- Each accepts
campaignandpage. A page contains 1,000 items.