How Do You Write a Title and Description So an AI Retrieves Your Page?

By putting into each page's title, description and headings the names, figures and specific terms that already appear in its body, using the words the person asking would use. In a study from Yonsei University and Konkuk University, accepted at KDD 2026, improving only the structural fields —title, meta description, headings and markup— made a page show up among the top 20 candidates of the retrieval stage in 71% of queries, against a 58% starting point. Rewriting only the body text left it at 53%.

It matters because before a model writes an answer, a search stage has to retrieve your page out of thousands. That stage compares the words of the question with the words in each field of the page, and the title, description and headings are short fields that summarize what the page is about. This article is the practical side of another one we already published, where we explain why optimizing only the body text can keep you from being found. Here we cover how to write those fields.

Why do the title and description weigh so much when a page is retrieved?

Because they are short fields meant to summarize what the page is about, and the retrieval stage rewards a match between the words of the question and the words of the page. The authors explain it this way: structural information is designed to concentrate terms relevant to the query, which increases the word overlap that retrievers prioritize.

In the system they built, each fragment of the body travels together with the title, description and headings of its page, and those fields are indexed separately. In other words: in that design, a vague title accompanies every fragment of the page.

The system is a test bed of their own: a word-based retriever, a reranker and a generator, over about 171,000 web pages from nine subject areas. It does not reproduce any particular commercial engine. What it did show is that rewriting only the body did not reliably improve retrieval, and often lowered the page's position, with word-based retrievers as well as with semantic and hybrid ones.

How much does each field contribute on its own?

The study does not separate them. It measures the set of title, meta description, headings and JSON-LD markup, which the authors treat as "structural information". That is why the 22% cannot be attributed to the title alone or the description alone. What it does allow us to say is that working that set with the terms already in the page's body gives the page better chances of being retrieved than leaving the fields vague or rewriting only the text.

What changes between a generic title and a specific one?

The specific one names the topic and its protagonists. The study gives two examples from its own tests. A generic title, "Panel Clarifies Advice", became "IOM Panel Clarifies Vitamin D Guidance", with terms specific to the subject. And a long, vague description became a short summary with concrete facts: year, genre, review source and a film's box-office gross.

Which words should you use?

The ones the person asking uses. Among the strategies the authors tested, those that replaced common words with technical or uncommon terms were among the ones that lowered retrieval the most: 14% less with technical terms and 8% less with uncommon words. One example from the study: swapping "eating" for "alimentary routines" or "sleeping" for "somnolence" reduces the overlap with the words of the question, and the retriever gives the page less relevance.

Stacking keywords does not work either. Aggarwal et al. (KDD 2024) measured that adding the query's keywords to a text barely improved a source's visibility. Writing with your customer's vocabulary is not repeating it: it is naming things the way the person looking for them names them.

Can you overdo optimization?

Yes, in three ways. First: touching everything at once pays less. When the authors optimized the structural fields and the body at the same time, the retrieval gain was 15%, lower than the 22% from optimizing only the fields, because the body changes took back part of what was gained.

Second: promising in the title what the page does not answer. In the ordering stage, the study observed that the reranker favored content that answered what the query was looking for and penalized content that stretched the topic beyond it. That is why the title and description should promise only what the page actually answers.

Third: rewriting what is already good. In the shopping area, the authors saw that every strategy they tested lowered the probability of being cited, and they attribute it to those pages already tending to be well organized for the buyer and to the queries in that area tending to be casual and recommendation-seeking. The staged method they propose includes the same criterion: leave fields that are already clear and specific as they are.

And the body text?

It is still what gets cited. The authors analyzed where the citations in the final answer come from and found that the vast majority come from the body text, not from the structural fields. The fields help you get found; the body helps you get cited. Their staged method combines both —reinforcing the fields with the body's key terms and opening the body with the main claim— and it was the strategy with the largest overall gain in their comparison: 28% more retrieval.

What can you do today?

• Title: say what the page is and what it covers, using the specific name (product, service, place) and not only the brand.
• Description: summarize in one or two sentences what the page answers, with two or three concrete facts already in the body: a figure, a year, a place, a model.
• Headings: write them as the question your customer would ask, and include the main term.
• Words: use your customer's. Do not swap them for rare terms to sound technical.
• Consistency: make the title, the description, the markup and the page text say the same thing. Google asks for this for its markup: it has to match the visible text.
• Leave what already works. Change only the vague fields.
• Start with the pages you most want to be found.

These are changes of a few lines per page, and they give each page better chances of being retrieved.