Map Extract

AI lead enrichment is only as trustworthy as its evidence. Ask the model to research missing fields only, require a source URL and an exact quoted passage for every value it proposes, check that the quote really appears on that page and that the page belongs to the right business, and leave the field blank when the evidence is missing. Unsourced AI fields look complete and are often wrong in ways you cannot see until a prospect replies.

This explainer is part of the search, website research and AI enrichment series. It covers how AI enrichment fails, how grounding with a URL and quote works, why missing-only research matters, and a review routine you can run every week.

How AI enrichment fails

Language models are built to produce plausible text. For lead data, plausible is the problem: a wrong value that looks right is worse than a blank, because nobody checks it.

Failure modeWhat it looks likeWhy it happensHow to catch it
Invented contact details[email protected] when the site only lists hello@The model fills a common patternRequire the exact address in a quote from the business's own site
Invented facts"Founded 2009", "12 employees"Answered from memory or guessedRequire a quote; treat size estimates as unsupported unless stated
Wrong entityDetails of a same-named company in another town or countryName collisionCheck the cited page's domain, address and company number
Wrong branchHead office phone for a local branch, or the reverseMulti-location businessesStore location with each value
Stale dataAn old address or a closed businessModel knowledge or cached pages are out of datePrefer live pages; record when the page was read
Citation launderingA real URL that does not contain the claimThe model cites a plausible pageFetch the page and match the quote
Directory as sourceValue taken from a listing site, not the businessEasy pages rank firstPrefer the business's own domain; label third-party sources
Overwriting good dataA captured phone number replaced by an AI guess"Fill all fields" promptsMissing-only research; never overwrite captured facts
Personal data creepDirector names, personal mobiles, LinkedIn profilesBroad promptsScope the fields; exclude personal data unless you have a reason and a lawful basis

None of these is visible in a spreadsheet. A column of tidy, fully populated values is not evidence of quality.

Grounding with a URL and a quote

Grounding means making the model answer from retrieved evidence rather than from memory. Google's Gemini documentation describes grounding with Google Search as connecting "the Gemini model to real-time web content", and says it can "reduce model hallucinations by basing responses on real-world information" while citations "build user trust by showing the sources for the model's claims". The same documentation describes combining search grounding with a URL context tool to ground answers in specific URLs you provide.

Citations alone are not enough for lead data. A citation tells you which page the model looked at; it does not prove the page says what the model claims. Add three things:

  1. An exact quote. For every value, the model must return the passage from the page that contains it, copied verbatim.
  2. A source URL per value, not one URL for the whole record.
  3. A confidence or support level, so weak support can be held rather than accepted.

Then validate outside the model:

CheckPass condition
Quote presentThe quote appears on the page at the URL (after normalising whitespace and case)
Value supportedThe value appears in the quote, or follows directly from it (for example a category from a services list)
Right businessThe page is on the business's domain, or a third-party page that names the business with matching address or company number
Field formatEmails pass syntax checks, UK phone numbers have a valid format, postcodes are valid
No overwriteThe field was empty, or the new value is clearly better supported than the captured one

A value that fails any check stays blank or goes to a review queue. It never enters the clean record because it "looks right".

An example request and response

Request a structured answer, with "not found" as an allowed outcome:

{
  "business": {"name": "Northfield Plumbing", "website": "https://northfield-plumbing.example"},
  "known": {"phone": "0113 496 0000", "town": "Leeds"},
  "research": ["services", "areas_served", "opening_hours", "general_enquiries_email"],
  "rules": "Only use pages about this business. Return value, source_url and an exact quote for each field. If not found, return null. Do not infer email addresses."
}

A good response (fictional):

{
  "services": {"value": ["boiler installation", "bathroom fitting", "landlord gas safety checks"],
               "source_url": "https://northfield-plumbing.example/services/",
               "quote": "We offer boiler installation, bathroom fitting and landlord gas safety checks"},
  "areas_served": {"value": ["Leeds", "Bradford", "Wakefield"],
                   "source_url": "https://northfield-plumbing.example/about/",
                   "quote": "covering Leeds, Bradford and Wakefield"},
  "opening_hours": {"value": "Mo-Fr 08:00-17:30",
                    "source_url": "https://northfield-plumbing.example/contact/",
                    "quote": "Monday to Friday, 8am to 5.30pm"},
  "general_enquiries_email": null
}

The email is null because no address appears on the site. That blank is correct and useful: it tells you to use the contact form or phone instead.

Use structured data where it exists

Many business sites publish Schema.org markup. The LocalBusiness type covers address, telephone, email, openingHours, areaServed, legalName and vatID, among others. When markup is present, read it deterministically first and use AI only for what the markup does not cover. Markup is the business's own structured statement and needs no interpretation. Our website data extractor reads that markup before it looks at page text.

Missing-only research

The safest enrichment asks for less. Missing-only research sends the model:

  • a compact identity: name, website, town and any registry number;
  • the fields you already have, so it does not repeat or contradict them;
  • the specific fields you want, and nothing else.

Why this matters:

  • It protects captured facts. A phone number read from the business's own contact page is better evidence than anything a model will find. Do not let a "complete this record" prompt replace it.
  • It reduces cost and noise. Fewer fields means shorter prompts and fewer chances to invent.
  • It makes review possible. A reviewer can check three proposed values with quotes far faster than twenty.
  • It keeps personal data out by default. If you do not ask for named people, you are less likely to receive them.

A useful rule of thumb: if a field can be captured deterministically (email addresses, phone numbers, social links, company numbers in the footer, structured data), capture it that way. Use AI for fields that need reading and judgement: services, business model, audience, category, a plain-English description, or distinctive facts such as brands stocked.

Fields that suit AI and fields that do not

FieldAI suitabilityNotes
Services and productsGoodUsually stated on the site; quote supports it
Category and nicheGood with evidenceClassification from the services quote
Areas servedGoodOften stated on about or service pages
Business description in plain EnglishGoodKeep the source text; label as a summary
Opening hoursGoodOften in markup or on the contact page
Email and phonePoorCapture deterministically; never infer
Company size, revenuePoorRarely stated; estimates are guesses
Founding yearFairOnly when stated ("since 2009")
Named decision makersAvoid by defaultPersonal data; needs a lawful basis and care
Technology usedPoor via promptDetect from the site's code instead; see finding businesses by website technology

A review routine that holds up

AI enrichment needs ongoing measurement, not a one-off test.

When you start a new niche or prompt:

  1. Run 50 records.
  2. Check every accepted value against its quote and page.
  3. Record each failure by type, using the failure table above.
  4. Adjust the prompt, the allowed fields or the validation rules.
  5. Repeat until failures are rare enough for your use.

Every week after that:

  1. Sample a fixed number of accepted values, spread across fields.
  2. Open each source URL and confirm the quote and the business.
  3. Review everything held as low confidence or inconclusive.
  4. Track the share of values that pass, by field. A falling rate for one field means a source changed or the prompt drifted.
  5. Re-run research when the source page changes; an old quote no longer supports a value the page has removed.

Before outreach, check the fields you will use in the message. If a personalised line says "I saw you now offer landlord gas safety checks", the quote should be on the record.

Data protection and providers

Whatever you put in a prompt goes to the provider you choose, under that provider's terms. Keep prompts to the business identity and the fields requested. If you research named individuals, UK GDPR applies: the ICO is clear that information identifying a named person is personal data even in a business context, so you need a lawful basis, minimisation and transparency. A local model avoids sending data to a third party but still needs the same checks on its output.

How Map Extract helps

Map Extract treats AI output as proposals that must be validated before they are written. In the Windows desktop app, AI research is optional and uses the provider you configure: the Gemini API by default, other API providers, a local model server, or a supported AI chat site in a browser profile.

  • Missing-only research. The live research route sends a compact business identity, the fields already held and the requested gaps; known links such as the website, sitemap or contact page guide the research. A separate route analyses evidence already captured from the site.
  • Proposals validated against source passages. Proposed values come back with a source and an exact quotation, and the application checks the source passage, whether it supports the field, and the link between the source and the business before writing eligible fields.
  • Unsupported values stay blank. An inconclusive result can be retried or held for review. Captured facts are not overwritten by AI.
  • Provenance stays visible. Records show the value, its source URL and the evidence passage where available.
  • Layer 3 extra facts. For distinctive details such as equipment models, brands stocked or published warranty terms, Layer 3 uses at least two model reviewers and requires each fact to be tied to an exact page URL and a quotation on a captured page. It excludes personal contacts and other personal data, and a certification listed on a site is reported as that site's claim.
  • Deterministic capture first. The website data extractor captures emails, phone numbers, social links, company numbers, structured data and technology signals from the pages themselves; technology signals come from capture, not from the research prompt.
  • Search-side AI. In the SERP scraper (also presented as the Google search scraper), AI can suggest keywords and domains to suppress for you to review. The AI business finder proposes niches and locations, and its suggestions never become business facts on their own. The workflow builder saves the source order so research runs the same way each time.

What it does not do: it does not guarantee that every field will be filled or that every accepted value is correct, and provider access, terms and charges depend on the service you configure. Review held records and sample accepted ones.

Public business and social results the app collects are uploaded automatically to Map Extract's servers while it is licensed and running; the privacy notice explains this.

Key takeaways

  • Unsourced AI fields look complete and fail silently; require a URL and an exact quote per value.
  • Validate outside the model: quote present, value supported, right business, valid format.
  • Ask for missing fields only, and never let AI overwrite captured facts.
  • Capture contacts deterministically; use AI for services, category, areas and descriptions.
  • Blank is a valid answer. Hold weak support for review.
  • Sample and measure accepted values every week, by field.

Frequently asked questions

Can I use ChatGPT or Gemini to enrich a lead list?

Yes, for research and classification, as long as every value it returns comes with a source URL and a quoted passage you can check. Without sources, a model will often produce plausible but invented details such as an info@ address, a phone number or a founding year. Ask for missing fields only, allow 'not found' as an answer, and verify quotes before you accept anything.

What is grounding in AI enrichment?

Grounding means tying the model's answer to retrieved evidence, such as a web page, instead of letting it answer from memory. Google's Gemini documentation says grounding with Google Search can reduce hallucinations and that citations show the sources for the model's claims. For lead data, go one step further: require an exact quote and check it appears on the cited page.

How do I stop AI making up email addresses?

Never accept an email address from a model unless the exact address appears in a quoted passage on a page that belongs to the business. Do not let the model infer addresses from patterns such as first.last@domain. Then run normal email checks for syntax and mail routing, keeping catch-all and inconclusive results labelled. Unsupported addresses should stay blank.

How accurate is AI lead enrichment?

It depends on the evidence, the field and the checks, so measure it yourself rather than trusting a vendor's figure. Review a sample of accepted values against their quotes each week and record why rejected values failed. Fields with clear evidence on the business's own site, such as services or opening hours, tend to hold up better than inferred fields such as company size.

Is it safe to send lead data to an AI provider?

Whatever you include in a prompt goes to the provider you chose, under its terms. Send only what the task needs: the business name, website and the fields you want researched. Avoid sending personal data about named individuals unless you have a lawful basis and the provider's terms suit it, and check whether the provider uses inputs for training.

Sources

  1. Grounding with Google Search (Gemini API docs)
  2. Gemini API documentation
  3. LocalBusiness (Schema.org)
  4. Business-to-business marketing (ICO)

Written by the Map Extract editorial team. Found an error? Email [email protected] and we will check and correct it. How we research guides: editorial policy.

Find B2B Leads With Google Search Operators

Operator recipes for finding B2B prospects in any niche, how to keep directories and marketplaces out, country and language settings and repeatable lists.

10 min read

Find Wholesalers and Distributors in Any Niche

Search recipes to find wholesalers and distributors in any niche, how to tell manufacturers, distributors and retailers apart, and what to check first.

8 min read

See it working on your own searches.

Start with one keyword, one location and a small limit. Review the records, then decide on a licence.

Top