Adapted from Duane Forrester's analysis on Duane Forrester Decodes. Aurora extends his framing for marketing teams running AI-search visibility programs.
Keyword research never lied to you. It was visibly blunt. You knew you were approximating relevance through lexical overlap, and that visible bluntness kept the practice honest. You over-covered topics, built supporting clusters, and triangulated intent precisely because you could see the instrument was crude.
Vector-based alignment scoring is a genuine upgrade. Embedding models capture semantic proximity that keyword tools cannot see. A page can use none of your target keywords and still be strongly aligned to a query because it covers the same conceptual territory through different vocabulary. That paraphrase space is structurally invisible to lexical analysis. The new instrument crosses a gap the old one could not.
It is also more dangerous, because precision is not accuracy.
A higher-resolution approximation is still an approximation
The vector space model dates back to Gerard Salton's SMART system at Cornell in the 1960s. The math has changed dramatically — modern transformer embeddings encode meaning in hundreds or thousands of dimensions — but the underlying move is identical. Represent the query and the document as vectors, measure the angle between them, treat that angle as a proxy for relevance. The proxy got sharper. It is still a proxy for a relationship that exists outside the math.
Netflix researchers Steck, Ekanadham, and Kallus demonstrated in 2024 that cosine similarity over learned embeddings can produce results they describe as arbitrary. How a model was trained, what data it saw, what regularization was applied — all of it shapes the geometry of the embedding space in ways that make a raw cosine score unreliable as an absolute measure of meaning. A 0.92 in one model is not a 0.92 in another.
For practitioners optimizing content, the consequence is direct. When you score your article against a query using an embedding model, you are measuring semantic proximity inside that specific model's representation of language. You are not measuring how OpenAI's RAG pipeline, Perplexity's Vespa-based retrieval, or Google's Gemini grounding would evaluate the same relationship. Those systems run their own embedding models, their own reranking, and their own filtering. A high score in your tool can map to weak retrieval in production.
Known unknowns vs. unknown unknowns
Keyword research produces a known unknown. You see the limitation. Vector alignment produces an unknown unknown — a number with decimal places that feels settled. That feeling is the trap. The score says nothing about whether the production system's embedding geometry resembles yours, nothing about how reranking will treat the result, nothing about whether the generation layer will deem your content authoritative enough to cite.
Goodhart's Law applies cleanly. The moment alignment becomes the target, the content drifts toward the geometry of the measurement model and away from the actual relevance it was meant to approximate. You start writing for an embedding model that no production system uses.
Where Aurora fits
This is exactly the gap Aurora is built to close. We don't ask whether your content scored high in an offline embedding tool. We watch what ChatGPT, Perplexity, Gemini, Google AI Mode, Copilot, and the rest actually do with real prompts, on a schedule, over time:
- Did your domain get cited? Aurora extracts citations from every tracked response and tallies share-of-citation per prompt, per engine, per competitor.
- Did your brand even get mentioned, with or without a link? Brand visibility scoring catches paraphrased mentions and bare references the citation count would miss.
- Is the gap a content problem or a discoverability problem? Looking at retrieval-grounded responses vs. memory-only responses tells you which lever to pull.
- Did a model update just shift the geometry? Adaptive querying surfaces the prompts where citation patterns changed week-over-week so you see drift before it shows up in traffic.
None of that replaces alignment scoring. It complements it. Use vector alignment to direct your editing; use Aurora to confirm whether production systems agree. Treat the alignment number as a directional signal — useful, not definitive — and the production signal as the ground truth that actually pays you back.
What good measurement literacy looks like in 2026
Three habits separate the practitioners who get value from these tools from the ones who optimize themselves into a corner.
- Pair the offline score with a production check. Never let an alignment number be the only thing you measure. Run the same prompt through real engines. If the engines disagree with your score, the engines are correct.
- Track the spread across engines. A 0.9 alignment that lifts you in ChatGPT and not in Gemini tells you something an aggregate score never will. Aurora's per-engine breakdown exists for this reason.
- Watch the geometry move. Embedding spaces and production reranking change. Treat alignment scores as time-series, not constants. The story is in the slope.
The gut feeling was never the enemy. The illusion that you have moved past the need for judgment is. Vector alignment is the most sophisticated approximation we have ever had. It is still an approximation, and the discipline of knowing what the number is not telling you is the entire job.
