Entity Extraction and Summarization
Extracting a page mechanically shows what it communicates to a machine, which is usually narrower than its author intended. Ivan Sp extracts the structured content of a page as an audit, listing every entity, relationship, figure, and call to action it contains. The output shows what a page actually communicates to a machine, which is frequently narrower than what its author intended.
The Four Extraction Categories
Extraction sorts a page’s content into four categories:
- Entities. The company, its locations, its products, and its services, each with the exact name the page uses.
- Relationships. Who provides what to whom, stated as explicit connections rather than implications.
- Numerical data. Founding year, years in operation, quantities, prices, and response times.
- Actionable elements. Every call to action, recorded with its exact button text.
One worked example ran against a security company’s key-card access page. The extraction returned the company and its location, the product and service relationships, the founding year and years in business, and each call to action verbatim.
What the Output Reveals
The value sits in what the extraction fails to find. A page whose author believes it explains three services, and whose extraction returns one, has a structural problem rather than a content problem.
Extraction also exposes claims made only through implication. A page implying a service area through photographs and local references, without ever stating the service area in text, extracts with no geography attached.
Checking Schema Against Reality
Extraction output gets compared against the page’s structured data. The two should describe the same business, the same services, and the same figures.
Disagreements get flagged. Schema declaring a service the visible content never mentions is the most common finding, and it breaks the rule that structured data must match what a visitor can actually see. That rule exists because mismatched markup triggers rich-result removal.
Summarization for Downstream Systems
A related build compresses a page into a structured summary that another AI system can consume. The format carries business identity, the primary offering, value propositions as complete sentences, solution mechanics, problem and solution pairs, customer segments, service tiers, categorized benefits, and every call to action.
Compression follows one hard rule. Any sentence containing a causal or conditional connector stays intact, because truncating at “because” or “when” destroys the claim. The same rule governs the business profile page written for answer engines.
Extract One Page Yourself
I suggest a manual version that takes about ten minutes.
- Copy the visible text of one page into a blank document.
- Highlight every proper noun, product name, standard, and place.
- Underline every number, date, and measurement.
- Count both, then compare the totals against what you thought the page covered.
Most authors expect three times what they find. The difference is everything the page implies and never states.
Where the Standard Content Review Advice Breaks Down
What would a machine list if it read your page with no prior knowledge?
- The usual advice
- Read the page carefully and judge whether it covers the topic well.
- What holds up in practice
- A careful human reader supplies missing context automatically, from the rest of the page and from their own knowledge. Mechanical extraction supplies nothing, which is why it finds the gaps a close reading cannot.
Consider an author reviewing their own page and finding it comprehensive. Every point they intended to make is present to them.
The extraction lists four entities and two relationships. Everything else the author believed the page said was implied rather than stated.
Ivan Sp extracts mechanically before judging. The gap between what a page implies and what it states is invisible to anyone who already knows the subject.
A page can read as complete to its author and communicate almost nothing to a system. The author is the one person who cannot detect that.
Everything your page implies rather than states is invisible to the systems describing you. The author is the last person who will notice.
Four Things Commonly Assumed About What a Page Says
Which of these do you assume your pages communicate?
I meet these four assumptions on most extraction audits.
| Commonly assumed | What actually holds |
|---|---|
| A careful reading finds the gaps. | A human reader supplies missing context from knowledge and from the rest of the page. Mechanical extraction supplies none. |
| Implied information counts. | Implication requires inference. Extraction records what is stated, which is the same subset a machine works from. |
| More words means more entities. | Length and entity density are unrelated. A long page can name very few specific things. |
| Schema markup describes the page. | Markup describes what somebody claimed. Extraction shows whether the visible text supports those claims. |
| The author knows what the page says. | The author knows what they meant. The two diverge most on pages the author knows best. |
The version of your page that machines read is shorter and vaguer than the one you wrote. Extraction is the only way to see it.
A Worked Example: What Extraction Returns
Consider a 900-word service page put through mechanical extraction. The output below is illustrative, written to show the typical gap. No client engagement is described here.
| Category | What the author expected | What extraction returned |
|---|---|---|
| Named entities | About 20 | 6 |
| Explicit relationships | About 15 | 4 |
| Figures and thresholds | Several | 1 |
| Unambiguous calls to action | 1 | 0, since the ask was implied |
| Schema claims supported by visible text | All | 3 of 5 |
| Terms used for the core concept | 1 | 3 |
I treat the right column as the brief. Each shortfall names something the page could state and currently only implies.
Reading your own page for gaps is proofreading your own writing. You read what you meant, every time, and the missing word stays invisible.
The schema row is the serious one. Two marked-up claims with no visible support is a structured data policy issue rather than a content gap.
How an Extraction Audit Runs
I extract first and judges afterwards, since judging first contaminates the extraction.
- Strip the page to its visible text, removing navigation, markup, and images.
- List every named entity, including products, companies, places, standards, and people.
- List every explicit relationship stated between those entities.
- List every figure, measurement, date, and threshold the text contains.
- Record the call to action exactly as worded, or record that none is stated.
- Compare the entity list against what the page was intended to cover.
- Compare the structured data claims against what the visible text actually supports.
- Flag every term used in more than one form for the same concept.
- Turn each shortfall into a specific addition rather than a general instruction to expand.
- Re-extract after the additions and confirm each gap closed.
How This Engagement Is Scoped
I scope extraction audits by page count, since each page is extracted individually.
- What arrives
- A four-category extraction per page, plus a specific addition list per shortfall.
- Typical duration
- One to two weeks for a set of twenty pages.
- What the client supplies
- The page list, and the facts needed to close any gap found.
- Included
- Extraction, comparison against intent, schema cross-check, and the addition list.
- Not included
- Writing the additions, quoted separately once the gaps are agreed.
- What changes the estimate
- Very long pages, or sites with heavy plugin-generated markup to reconcile.
- How it is quoted
- After a short scoping call, once the page list is agreed.
Tools and Primary Sources
The references below cover the vocabulary and the policy. Each opens in a new tab.
- Schema.org vocabulary supplies the entity types and properties an extraction is compared against.
- Google structured data policies requires markup to describe content visible on the page.
- Google AI features documentation describes how AI surfaces draw on the stated content of a page.
- Schema Markup Validator checks the markup side of the comparison against the full vocabulary.
Common Questions About Entity Extraction and Summarization
How is extraction different from a normal content audit?
A content audit judges quality, tone, and completeness for a reader. Extraction records only what a machine can identify without interpretation, which makes the gaps visible in a way a readability review does not.
How many pages get extracted in an engagement?
Extraction runs by hand on high-value pages such as the homepage and primary service pages. Site-wide extraction runs through the crawl pipeline instead, which scores every page at once.
What gets done with the extraction output?
The output drives three things: schema corrections where markup and content disagree, content additions where an intended claim never got stated, and the source material for a business profile page.
Why not just read the page carefully instead?
A careful reader fills gaps from their own knowledge without noticing. Mechanical extraction has no knowledge to fill from, which is exactly why it finds what a close reading misses.
Discuss This Work
The first call is a free 30-minute consultation with no pitch attached. You receive an honest assessment of where the site stands and what can realistically be improved.