Find the business.
Build the evidence.
Make the next move clearer.
A practical account of agentic prospection: choosing the search method, defining the audience, retaining website evidence and qualifying social accounts at scale.
Same target. A different way to get there.
The reported trials targeted 300 businesses. Changing how the agent collected them reduced elapsed collection time, while people still defined the scope and judged the result.
Trial timings reported by the project owner. These are elapsed collection times, not a controlled human-versus-agent benchmark or a guarantee of identical output quality. Five concurrent runs are a proposed trial: 90 ÷ 5 gives 18 minutes with perfect scaling. Around 15–20 minutes is the target; rate limits, retries and consolidation may add time. Plan capacity still needs checking before that run.
The human work
Choose the market, define useful categories, agree the evidence and approve the boundaries. Investigate exceptions and decide what the findings mean for a business.
The agentic work
Repeat the searches, capture sources, reconcile records and route gaps into the next collection step. Scripts make the same checks repeatable across the batch.
Scope first. Then release the work.
The Prospection Time workflow defined the town, exact search geography, catalogue and collection method: agent-led scraping, Firecrawl, or two concurrent Firecrawl runs. ONS population supplied town-size context and helped set the collection budget. A recorded queue and approval step kept the chosen scope attached to the work.
Which searches returned results?
Compare every keyword across the towns in the acquisition workbook. A number is the candidate count returned by that query; Failed is a recorded query failure.
Return rate = queries with at least one candidate ÷ recorded queries. This measures query returns, not successful website scrapes. Candidates may repeat across keywords. A dash means no record; Bovingdon is absent from the B and C tabs.
Workbook snapshot · 11 September 2026. Watford built-up area in C is displayed under Watford. View the query-performance source ↗
Three catalogues. Different working hypotheses.
A, B and C grouped professions before collection. The hypothesis was that different kinds of business would have different operational gaps and different opportunities for automation.
Trades, automotive, personal care and beauty
21.3%without an own-business website recorded
722 of 3,391 matched businesses
Hospitality and events, transport and hire, facilities, industrial supply and repair
16.3%without an own-business website recorded
439 of 2,697 matched businesses
Clinics and healthcare, property, business and professional services
12.2%without an own-business website recorded
361 of 2,968 matched businesses
That first comparison supports one part of the hypothesis: the recorded website gap is larger in A than in C. It does not establish how much automation a business needs. A working website can coexist with slow enquiries, manual administration or an inactive social presence.
How this comparison was calculated
Matched the retained business records’ profession labels against the approved A, B and C catalogues: 41, 40 and 40 labels. Each business is counted once within each matched group. Businesses can match more than one group, and these groups do not cover every retained record, so the totals should not be added together.
“Without a website recorded” means the main record is not classified as “Business website”. Directory and high-street entries do not count as the business’s own website. This is the same main-record basis used by the explorer; it does not include the separate follow-up search for website links on social profiles.
Capture once. Keep using the evidence.
Direct HTTP scripts retained website HTML without provider credits. The saved page could then support contact extraction, service information, social-link discovery and later assessment without fetching it again.
domains with HTML retained
From 2,893 attempted domains in the documented August campaign. This is one campaign, not the full study’s capture total.
raw social links found
Reconciliation reduced repeated links and URL variants to 4,008 canonical account candidates on 1,625 businesses in that campaign.
The fallback route handled different gaps. Exa searched for missing website destinations; candidate verification checked the match. Firecrawl captured difficult sites that ordinary retrieval had not resolved. The named difficult-site batch retained pages from 559 of 607 domains.
A captured page was not automatically accepted as the correct business website. The round closed with reconciled results and an explicit unresolved tail, rather than treating every failed retrieval as proof that no website existed. Zero provider credits applies to the direct lane, not to labour or infrastructure.
Turn a profile link into usable evidence.
Make and Apify workflows, together with X and YouTube APIs, supplied profile and publishing information that could be linked back to the business record.
Collect and connect
Account identity, biographies, contact fields, followers or subscribers, dated posts and available responses. Reconcile duplicate account links and distinguish readable own-business accounts from shared, unsuitable or unresolved candidates.
Compare consistently
Use the same reference date for recency, compare audience size within each platform and keep missing dates separate. These are retained snapshots: newest-post recency is not a measure of posting frequency or audience growth.
A search result is just the beginning.
Three phases connected the business identity, its website and the evidence on its social accounts.
Find the business.
Keep the source.
Profession and location searches supplied the starting list. Records were organised and consolidated so the same business could be followed into the next phase.
17 fields
The original round handed over 4,703 business records with names, locations, links, contact details and source references.
Turn a destination
into useful detail.
Website capture supplied the content, contact routes, technical features and social links behind each assessment. The original round expanded those same 4,703 records to 60 fields.
60 fields
Connect the profile
to the work.
Account detail and dated posts added audiences, descriptions, contact information and activity. Website and social evidence could then be read together for the same business.
272,063
Populated cells in the current social-account export, across 53 fields.
Meet the tools behind the research ↗Business identity
Illustrative record structure · no individual business data
Different scores answer different questions.
The website quality score and the wider qualification scores are separate measures. Their scales and inputs need to stay visible.
Website quality
The explorer’s recorded website assessment. The original model used weighted checks for HTTPS and mobile readiness (2 each), structured data and an enquiry form (1.5 each), analytics, modern image formats and content depth (1 each).
The score describes detected website features. It is not a Google review rating, a full accessibility audit or a measure of business performance.
Combined qualification
The wider rubric brings together website qualification /16, social qualification /14 and business qualification /6. Earned points across the 36 factors are normalised to 100.
This is the stored Agentic score. It assesses recorded digital and contact evidence, not a proven level of AI readiness or the value of an automation project.
Website factors
Own-site presence, security, mobile and content signals, contact and social links, title, description, heading, indexability, video and explicit motion.
Social factors
Six own readable platform accounts, activity within 30 days, response on the newest post, audience against its platform median and profile completeness.
Business factors
Google listing, graded star rating, at least ten reviews, email, phone and a matching website across sources.
What the scores can—and cannot—tell us
Most qualification factors earn one point when their rule is met. Google stars use a graded point: 5.0 earns 1; 4.5–4.9 earns 0.75; 4.0–4.4 earns 0.5; below 4.0 or unrated earns 0.
Several rules award no point when evidence is not found. That can reflect missing collection as well as a real business gap. Read the score alongside evidence coverage, account ownership and capture dates. HTML checks are proxies: a viewport tag does not prove a good mobile experience, and a detected form may be a newsletter form.
The qualification rubric’s 30-day activity check is distinct from the explorer’s 90-day comparison. Historical social scores are retained separately and are not substituted for the current qualification /14.
A repeatable process, with the evidence still attached.
Define the search. Capture and reconcile. Enrich the record. Qualify the evidence. Then use the results to choose a practical next step.
Explore your profession →Sources and scope
Approved search-led A and Places B/C catalogues, version 1.0.0; retained combined business view; August qualification round closeout; current qualification evaluator 0.6 and its 36-factor rubric. Catalogue comparison calculated from the retained 8,976-record study. Timings are owner-reported observations; the five-run estimate has not been tested.