A technical article on extracting fields from wedding photography websites points out that missing fields take at least three forms: the information exists in an image or PDF, multiple pages conflict, or the information never existed and the merchant explicitly says they do not offer it. We note that lumping all three cases under “missing” leaves subsequent data-remediation work without direction: the first two can be fixed, while the third requires new business rules or abandoning the question.
What this is
The core approach is to replace boolean judgments with multi-state classification. The original example returns states such as “present in text,” “present in an image,” “conflicting statements,” “explicitly not offered,” and “genuinely missing.” It uses regular expressions (expressions that match text rules) to inspect page content, then searches for clues through image alt text and nearby text.
Pricing images cannot simply be converted into text with OCR (recognizing text in images). They also require parsing of price ranges, package tiers, and number formats. Otherwise, “starting at RMB 6,800” and “starting at 6800 yuan” may be treated as different prices, creating false positives; the article says that without normalizing formats, the false-positive rate for conflicts may exceed 30 percent.
Another critical step is cross-page verification. If the homepage says “covering Wuhu and surrounding areas” while the pricing page says “city center only,” a single-page tool will not detect the problem, but comparing multiple pages will. Detection systems that crawl only the homepage cannot cover these easily overlooked differences in business terms.
Industry view
We believe the article’s value lies not in code details but in making a common mistake in enterprise AI projects concrete: data extraction is not a mechanical task of “found it or did not find it.” It requires judgment about provenance, conflicts, and business boundaries. For industries such as wedding photography, foreign trade, and manufacturing that depend on websites, quotations, and PDFs, these distinctions directly determine whether business teams can adopt AI answers.
The risks are also clear. Three-state classification depends on page samples, field rules, and semantic lexicons. Rules that are too narrow may misclassify new business lines as “not offered,” while rules that are too broad generate more conflicts and false positives. The original implementation still has limitations: some negative judgments depend on page conditions, and its OCR price parsing covers only a limited range of formats. It cannot be treated as a universal solution.
Moreover, “explicitly not offered” is not necessarily a problem. Merchants that proactively state “no out-of-town destination photography” or “makeup service not included” provide boundaries that AI can cite, reducing the risk that customers discover a service mismatch only at the deal-closing stage. What warrants closer attention is treating “not stated” as “available,” disguising incomplete information as a business commitment.
Impact on regular people
For enterprise IT: As companies introduce AI, IT departments need to inspect websites, PDFs, images, and different page versions—not merely test whether the model can retrieve keywords. The focus of data governance will shift from “filling empty fields” to “identifying conflicts and defining boundaries.”
For individual professionals: When using AI to research products, services, or suppliers, answers may come from outdated pages or materials from different versions. Before making an important decision, cross-checking the official website, quotation, and contract terms remains safer than relying directly on a single generated result.
For the consumer market: Merchants may increasingly publish what they do not support, much like they label their service coverage. This will make it easier for consumers to assess fit in advance, but if platforms extract such information incorrectly, gaps may still emerge in which a page says a service is available while the merchant actually declines it.