Crawling & Data Extraction

AI data extraction from web pages

Web pages being captured, checked, and converted into structured records
Web pages being captured, checked, and converted into structured records

Extraction should begin with a decision, not a crawler: what record do you need at the end?

"Extract useful information" is too vague to validate. A workable request names the fields, the permitted sources, how missing values should be represented, and which source URL supports each record.

Define the output before collecting pages

For a public supplier directory, the record might be:

FieldTypeRule
supplierNameTextUse the displayed legal or trading name
categoryControlled textMap to the agreed category list
locationTextPreserve the locality as published
certificationListInclude only explicitly stated certificates
sourceUrlURLKeep the page used for the record
capturedAtDate and timeRecord when the source was captured

Do not ask a model to infer a certification or location that the page does not state. Use an empty value and a review flag instead.

Capture a controlled source set

In Seekdown, a website data capture job starts from selected URLs and follows links within its hostname, path, depth, and page limits. Use included paths for the relevant directory and exclude navigation that produces duplicates or unrelated records.

Seekdown crawler controls for initial URLs and allowed hostnames with sample data
Seekdown crawler controls for initial URLs and allowed hostnames with sample data

Run a small capture first. Ten representative pages will reveal template variations and missing fields faster than a full-domain run.

Inspect the captured records

Open the memory dataset and search for several known product names, headings, or values. Check that the body contains the text needed for extraction and that each record retains its source URL.

Seekdown dataset contents used to inspect captured records with sample data
Seekdown dataset contents used to inspect captured records with sample data

If the important value appears only after a login, inside an image, or in a script the capture does not retrieve, change the source method rather than filling the gap by guessing.

Validate structured results

Build a review sample that includes common, sparse, and unusual pages.

CheckPass condition
CompletenessRequired fields are present or explicitly marked missing
AccuracyValues match the visible source
ConsistencyEquivalent values use the same format
ProvenanceEvery record retains its source URL
FreshnessThe capture timestamp fits the business need
ExceptionsPages with a different template are flagged

Calculate accuracy only from a labelled sample and record its size. For example, "47 of 50 reviewed records matched the source" is auditable; "highly accurate extraction" is not.

Export without losing provenance

Seekdown datasets can be exported as CSV or JSON. Enterprise datasets can also be exposed through an OData feed for Excel Power Query or Power BI. CSV exports include fields such as key, title, status, URL, body, summary, tags, source, and discovery timestamp.

Keep the source URL and capture time in downstream records. Without them, a reviewer cannot tell whether a surprising value came from the page, the extraction rule, or an old capture.