Why does cost of retrieval matter to AI models?
A model building an answer processes many sources at once. A page where the answer has to be dug out of ten paragraphs of running text is more expensive to handle than a competitor who put the same information in a table. With comparable content, the cheaper one wins.
The same mechanism works on the human side - structure that helps a reader scan the page helps a model find a fragment to quote. This is one of the few dimensions where optimizing for AI and optimizing for people is exactly the same work.
What will you see in the report?
You get a checklist marking which structural elements you have and which are missing. Every unmet point is a ready fix - usually a matter of minutes, because it needs no new content, only reorganizing what is already there.
Plus detected heading hierarchy defects: a missing H1, several H1s, or H3 sections with no parent H2.
Sample recommendations
Fragments of a report from an audit of an electronics store category page.
Problem: Parameters are scattered across a list instead of being laid out in a table — the model assembles the answer from four separate bullets.
Key technical parameters: 800–26,000 DPI resolution, acceleration up to 50G, 1 ms response time, new-generation optical sensors.
Add a table: Game genre | Recommended sensor / DPI | Connection type | Weight.
Problem: Empty H3 headings in the footer raise the processing cost — the parser walks a structure that holds no content.
### Customer service ### Shopping ### Current promotions ### Bestsellers ### Trending
Remove the headings from footer navigation or give each one a sentence of content; leave plain links as a normal list.
How do we measure cost of retrieval?
This is a points checklist, not a judgement call. We check for the presence of structural elements and add up the points to a maximum of ten:
What the score is made of
| Structural element | Points |
|---|---|
| Correct H1 → H2 → H3 hierarchy | 2 |
| Tables with data | 2 |
| Bulleted or numbered lists | 1 |
| Key facts in bold | 1 |
| A summary or TL;DR section | 1 |
| No walls of text over 300 words without formatting | 1 |
| Internal links embedded in context | 1 |
| No vague introductions | 1 |
What raises and what lowers the score?
Raises
- One H1, logical H2s beneath it, H3s only under those
- Numeric data presented as a table instead of listed in a sentence
- Lists wherever the content is an enumeration: steps, conditions, requirements
- Key facts in bold - but facts only, not whole sentences
- A short summary at the start or end of the page
Lowers
- An H3 with no parent H2, or several H1s on one page
- Paragraphs over 300 words with no subheading, list or emphasis
- Data scattered across sentences instead of collected in a table
- A vague introduction before the first concrete piece of information
- Bold scattered across the whole text - it stops meaning anything
Frequently asked questions
How is this different from Effort Score?
Cost of Retrieval measures whether the structure can be scanned quickly; Effort Score measures whether the page is complete. The first looks at hierarchy, tables and emphasis; the second at length against competitors, visual material and the update date.
Does bold text help?
Yes, but only on facts. A bolded key parameter or number works as a signpost; whole sentences bolded in every paragraph stop highlighting anything.
Does every page need a table?
No - a table counts when you have data to line up. If you compare parameters, prices or variants, a table lowers the cost of reaching the answer more than any other change. A text with no comparative data gains nothing from one.