Why generic AI outreach gets ignored now
Between 2024 and 2025, AI made it nearly free to generate ten thousand "personalized" cold emails, so everyone did. The result was predictable: inboxes filled with messages that swapped in a first name and a company name but said nothing the sender could not have written without ever looking at the business.
Recipients learned to spot the pattern in two seconds. Reply rates on template blasts dropped below 1%, and inbox providers responded with the February 2024 bulk-sender rules from Google and Yahoo, which made authentication mandatory and put hard spam-rate thresholds on senders. Agencies that blasted 5,000 generic emails a month started burning sending domains in three to six weeks.
The lesson is not "cold email is dead." It is that the cost of writing a bad email used to be your time, and now it is your domain reputation. The economics flipped: fewer, better emails now outperform volume on every metric that matters.
The research-first alternative
Research-first outreach inverts the usual order. Instead of writing a template and merging in names, you study each prospect first and let what you find dictate the email. If you cannot find anything specific to say, you do not send.
This is what Furet automates end to end: it sources businesses from Google Maps, fetches each prospect's website (homepage plus subpages), and runs an LLM over the actual content to find concrete gaps. Every email it composes must reference at least two real findings from that analysis, and each draft is scored 0 to 1 against a template-detection rubric. Only drafts above 0.80 auto-send; the rest are recomposed later or held for your review.
The principle works whether you automate it or do it by hand: every email references something real, or it does not go out.
What to actually research on a prospect's website
For an agency, the best findings are problems you can sell a fix for. You are not looking for compliments to pay; you are looking for billable gaps. The reliable ones:
- Missing SSL or mixed-content warnings. Still common on local-business sites, trivially verifiable, and an easy first project.
- A stale blog.A "News" page last updated in 2022 signals an abandoned content effort, which is an opening for an SEO or content retainer.
- No booking flow. A service business whose only conversion path is a contact form (or a phone number in the footer) is losing after-hours leads every week.
- Poor mobile rendering. Overlapping elements, tiny tap targets, or a desktop-only menu on a business that gets most of its traffic from Google Maps on phones.
- Missing schema markup. No LocalBusiness or Review structured data means weaker map-pack and rich-result presence, which is exactly what an SEO agency fixes.
- Slow pages. A homepage that takes five-plus seconds to load is a concrete, measurable problem with a concrete, measurable fix.
Two of these in one email is enough. A prospect who reads "your booking is phone-only and your blog stopped in March 2023" knows you actually looked.
Define a tight ICP from Google Maps categories and geography
Google Maps is an underrated prospect database for local and SMB outreach: every business has a category, a location, and usually a website. The combination of one category and one geography is a workable ICP.
Three rules make it tight instead of vague:
- Pick categories where website quality varies widely.Dentists, law firms, home-services contractors, gyms. If everyone in the category already has a polished site, your findings get thin.
- Pick owner-operated businesses. Franchises route your email to a head office that does not control the website. Independent operators read their own inbox.
- Size geography to your sending capacity. At 50 sends a day, one mid-sized metro area per campaign is plenty. Do not source 5,000 prospects you cannot reach this quarter.
Realistic funnel math
Here is what a research-first campaign actually yields, with conservative numbers:
- 1,000 businesses sourced from a category and geography.
- 600 to 700 with findable, verifiable emails. Email discovery plus verification (Furet uses the Reacher engine, with role accounts like info@ gated by default) typically clears 60-70% of prospects. The rest bounce-risk out before you ever send.
- 2-6% reply rate on personalized sends, versus under 1% for generic templates. On 650 sends that is 13 to 39 replies.
- Roughly a third of replies are interested. Call it 4 to 13 real conversations from 1,000 sourced prospects.
Compare that with a 1,000-send generic blast: under 10 replies, most of them annoyed, plus measurable damage to your domain reputation. Same list, very different outcome.
Step-by-step setup
- Pick one ICP: a single Google Maps category in a single metro. Resist the urge to run three at once.
- Set up a separate sending domain with SPF, DKIM, and DMARC, and let it warm up before real volume. Never send cold from your primary agency domain.
- Source a small first batch of 100 prospects and read the research output before anything sends. Check that the findings are ones you could actually sell against.
- Let the quality gate do its job. If a draft scores below threshold, it should be recomposed or held, not sent anyway to hit a number.
- Cap follow-ups at three total touchesper prospect: the initial email plus two short follow-ups, each referencing a new finding rather than "bumping this to the top of your inbox."
- Triage replies daily. With replies auto-classified (interested, not interested, out of office, bounce, auto-reply, question), the only ones needing fast human attention are interested and question.
- Review weekly and adjust the ICP, not the template. If replies are weak, the fix is usually a better-fitting category or geography, not cleverer copy.
You can test the whole loop without committing budget: Furet's free tier includes 10 researched prospects and 5 sends, enough to see what the research output looks like on your own ICP. Start with the free tier and read the drafts before you scale anything.