iSearchFrom

Together with one of the stakeholders, I tested whether users of the free URL modifier would pay for a real-IP search that shows true local results and ads. We mined specialist SEO communities for frustrations with existing tools, then interviewed working marketers.

Demand for an accurate proxy search was real. The bigger finding was a shift nobody had asked us about: marketers were seeing traffic arrive from ChatGPT, Gemini and other assistants, and had no way to see how those assistants talked about their brand.

That one insight widened the product. Alongside a better Google proxy, we committed to tracking AI answers the same way marketers already tracked rankings and ads.

In December 2025 we shipped a paid MVP built around the new Search Engine Proxy and retired the free tool. Agencies that needed local ad data adopted it quickly. Long-time users filled the inbox.

Instead of defending the decision, we asked them how the old tool fit into their day. The answer was speed: one query, one country, no login. We quickly brought back a modernised free search and the complaints died down. That lesson shaped everything after it: the free path is not a leak in the funnel, it is the front door.

My first modernised free tool (left) carried too much of the paid product's structure and felt like a demo of something you could not have.

On the new site, free search sits in the landing hero (right): one card with country, city and language, and a clear second path to the real-IP search when accuracy matters.

The MVP proxy (left) proved people would pay for real local results, but the form and the results competed for the same space.

In the shipped version (right), configuration lives in a fixed left rail and the result gets the rest of the screen, rendered inside the device it was searched from. Searches are saved to history, shareable by link and comparable side by side.

I built and maintain the shared UI package the platform is made from, and wrote most of its commits (136 of 177). The app, the marketing site and the client portals are all built from it.

• Type and colour: Geist and Geist Mono on warm neutrals, with one orange accent reserved for primary actions.

• Charts: a 10-colour data palette validated for colour-blind safety and contrast. Series always take colours in the same order, because the order is what keeps neighbouring lines distinguishable.

• Themes and languages: every screen works in light and dark, in English and Brazilian Portuguese.

The new platform launched on August 4–5, 2026, with sign-in moved to app.isearchfrom.com and existing accounts migrated. Since then we have shipped something new almost every week: Maps Visibility, an AI Data Chat that answers questions about your own data with charts, Amazon, Walmart and Google Shopping tracking, team access control and an MCP server for AI agents.

Every tool answers one question first and offers detail second. The dashboard leads with the four numbers a client asks about (top 10 keywords, ad share, local presence and AI mention rate), and each one links to the evidence behind it.

I rebuilt the first session around one goal: get every new account to a real result about their own site as fast as possible.

• Welcome screen: I explored five art directions side by side in a lab page, then retired the illustrations for a scannable list of tools. Clearer beat prettier.

• First campaign as an onboarding step: setup opens with the signup domain filled in and measures from the user's own country, so the first result is about them.

• Trial offer: the offer opens with the user's real rank, partly hidden, so the value is concrete before they decide.

Early relaunch data showed visitors picking the free search roughly three to one over the real-IP demo, even though the demo was visually promoted. Trying the real-IP search is what leads people to sign up and stay.

I set up a test with a single variable: two parallel buttons (control) against two radio cards with one submit button and the real-IP option preselected. Copy stayed identical in both arms. The radio layout roughly doubled the share of hero users who tried the real-IP search: 68% against 34.8%, across about 23,800 sessions. It became the default.

I also own the marketing site. A performance pass replaced the 10,000-dot world map behind the hero with a rasterised mask: the worst-case Safari frame rate went from 15 to 59 fps, the longest frame from about a second to 33 ms, and the home page HTML shrank by 84%.

After the relaunch I tracked search performance closely, restored a utility-first homepage title when branded click-through dipped, and added country landing pages for people searching for Google from a specific place.

Agencies do not just track search visibility, they explain it to clients every month. Client portals turn the data an agency already tracks into a branded report page their client opens, a PDF and a monthly email, all in the agency's name and all drawn from live data.

I designed and built the whole feature: the editor in the app, the client-facing portal and its library of 57 report blocks across rankings, ads, Maps, AI answers and store shelves. It is the largest piece of work I have shipped at iSearchFrom, and it is in final QA ahead of release on the Agency plan.

• Start in seconds: "Build it for me" picks blocks from what the agency already tracks for that client and writes every answer line. Or start from one of six goal templates.

• Phone order: the desktop grid flows to phone automatically, and any block can be reordered or hidden on phone without touching the desktop layout.

• White label: the agency's logo and colour with contrast checked automatically, a custom domain, and an optional "Powered by iSearchFrom" line.

• Competitor names hidden: rivals can appear as Competitor A, B and C, so a report never names other businesses to a client.

Shipping this much with two people only works if quality keeps up. Before release I ran an agent-driven audit across every portal block, state and screen size, with a second pass of agents challenging each finding. It produced 184 confirmed issues, 7 of them high severity, and 173 were fixed in the same pass.

Among the catches: every quoted AI answer labelled "Named 1st", and a rival with a similar name shown as "You". Exactly the kind of quiet error that would cost an agency its client's trust. The portal also ships with 814 automated tests of its own.

1. The free path is part of the product. The backlash taught us more about our users than any interview. Paying customers start as free users who trusted us first.

2. Design the failure states first. In a monitoring product missing data is normal. When it is designed, people trust the numbers that are there.

3. Test instead of debating. One clean experiment settled a question that opinions could not.

4. AI moves the bottleneck, it does not remove it. Agents get a screen to 80% fast. The last 20% (hierarchy, copy, edge cases) is still design work, and now it is most of the job.