Andrew Dunn

2023-2026

The Lifetime Value Co.

Senior Product Manager, SEO. Consumer data and people search.

verified
  • Product
  • SEO
  • AI systems
  • Experimentation

Context

The portfolio was 10+ brandsSource note 3 of consumer data brands covering people search, property data, and vehicle history, running at 90M+ annual sessionsSource note 1 and 1.9M+ ranking keywordsSource note 2. Traffic at that scale was already there. What was missing was a way to decide what to build next, and a way to know whether AI search was starting to take the traffic.

Two problems sat underneath that. Page templates were shipped on judgement rather than on evidence, so nobody could say which layout earned the conversion. And the answer engines had begun returning citations instead of links, with no instrument pointed at them.

Owned

  • Organic product strategy across the portfolio, and the AI and automation roadmap.
  • A cross-functional 8-person dev teamSource note 4, working from specifications I wrote in Jira from epic down to ticket.
  • Decision rights on template experiments, content investment, and what the monitoring system measured.
  • Roadmaps specified by hand. Execution delegated to the team, which is the split that made the Bumper roadmap survive past my tenure. That story is on its own page.

Architecture

The work that matters here is the LLM brand-monitoring product, because it is the piece that had no precedent to copy.

The measurement problem is sampling. Ask an answer engine the same question twice and it may cite different sources, so a single pass tells you nothing about whether you are cited. The system runs 10,000+ per querySource note 5, which is the volume that gets citation share to statistical significance rather than to anecdote. Below that, a shift in the numbers cannot be distinguished from sampling noise.

Each pass records citation share, sentiment, and ranking across ChatGPT, Claude, Gemini, and Perplexity, then writes to a warehouse so the series can be compared over time rather than read as a snapshot.

LLM brand-monitoring measurement loopA prompt set is sampled over many passes across four answer engines. Each pass records citation share, sentiment, and ranking. Results are written to a warehouse as a time series, which feeds the structured data and content decisions that change what the engines see on the next pass, closing the loop back to the prompt set.Prompt setSamplingN passesChatGPTClaudeGeminiPerplexityCitation shareSentiment, rankWarehousetime seriesStructured data andcontent decisions
Sampling is the step that carries the method. One pass cannot separate a real change in citation share from sampling noise.

The programmatic side used structured-output agents to track and generate pages across millions of indexable URLs, with quality gates in front of publication. Structured output matters here: an agent returning free text needs a human to read it, while an agent returning a typed object can be gated automatically.

Decisions

  • Statistical significance over dashboard speed. That prompt volume is expensive per run and slow to return. A cheaper sample would have shipped sooner and measured nothing. Rejected the fast version.
  • Structured output over free-text generation. It bounds what an agent can return, which is what makes an automated quality gate possible at all.
  • Experiment before template rollout. Every product-page change went through the experimentation program rather than shipping on judgement.
  • Roadmap authorship separated from execution. I specified, the team built. Slower to start, and the reason the work continued without me.
  • Audited before commissioning. Content spend went to a scored inventory rather than to new production, which is the cheaper order of operations.

Outcome

The experimentation program ran across 50M+ product-page splitsSource note 6, producing 5%+ conversion liftSource note 7 and 6%+ on tested pagesSource note 8, with 16% period-over-periodSource note 9 at the top of the funnel.

The content audit covered 1,300+ across 4 sitesSource note 10, scored on traffic, bounce rate, product connection, and topical authority, which turned an argument about which pages deserved investment into a sorted list.

On head terms, "reverse phone lookup" reached top 10, 87 KD, 800K+ monthlySource note 11, one of the portfolio's highest-volume terms.

The monitoring product has no traffic outcome to report, and inventing one would miss the point. It was an instrument, and its value was that decisions about citation and structured data stopped being guesses.

Revision

I would have built the monitoring system before the programmatic tracking rather than alongside it. The programmatic work assumed the traffic model that answer engines were in the middle of changing, and measuring first would have re-ordered the roadmap. I would also have put the audit scoring in front of a wider group sooner, because a scoring framework nobody argues with is usually one nobody has read.

Brands

The brands this work covered, across the portfolio.

People search
Phone
Property
Vehicle history and fleet
Consumer tools