Adapted from Duane Forrester's essay on Duane Forrester Decodes. Aurora extends his framing for marketing teams running AI-search visibility programs.
Duane Forrester started his career in SEO during the early online-gambling era, when the rulebook was being written in real time. The principle he carried out of that decade is the one we keep coming back to in 2026: without data, every decision is a guess, and guessing is expensive. AI visibility is now the discipline where that gap hurts the most. Platforms are forming opinions about your brand. They cite some sources and ignore others. And most teams have no reliable way to see what is actually happening before it shows up in revenue.
That is the gap Aurora was built to close — and it is the same problem Forrester identified when he was running Bing Webmaster Tools through its second-generation rebuild. Webmasters lacked any reliable view into how Bing was handling their sites. Opening that data turned out to be one of the largest drivers of new Bing Advertising accounts and contributed to Microsoft's search business turning profitable. The lesson generalizes: if you give operators evidence they can act on, they act. Withhold it and they stop trying.
The four states of AI visibility, made explicit
Forrester's most useful contribution is the framing that "AI mentioned your brand" is not one outcome but several. From a single tracked prompt, an LLM can be in any of four states with respect to your content:
- Knows your brand from training data and answers from memory, with or without grounding it in your live site.
- Retrieves your content live via its connected search index and uses it to construct the answer.
- Knows neither layer and constructs an answer from competitors or third parties.
- Combines both layers — memory plus retrieval — producing the most stable kind of brand presence.
Each state demands a different fix. A retrieval problem is not a memory problem. A memory problem is not a content-quality problem. Treating "AI visibility" as one metric collapses these into a single number that hides exactly the diagnosis you need.
Forrester reports an anonymized beta finding from a 100-response audit: 57 paraphrased matches to the target site's content, zero linked citations, zero direct mentions. That is an invisible gap. The model is reading and reusing the content without crediting it, and most monitoring stacks would report visibility as zero when it is actually high — in the worst possible way.
What this looks like inside Aurora
Aurora is structured around the same principle: AI visibility resolves into multiple signals, and the signals must be measured separately before they can be combined.
- Citation tracking. Aurora extracts cited URLs from every tracked engine response and computes share-of-citation per prompt, engine, and competitor. This catches the literal "did your domain get linked" question.
- Brand visibility scoring. Independent of citation, Aurora detects mentions, paraphrases, and aliases in the response text. This is the only way to see the 57-paraphrases-zero-links case.
- Source intelligence. Aurora rolls up the domains AI engines cite when answering category-level questions. That tells you whether the model is reaching for Reddit, your own site, a competitor, a news outlet, or an aggregator — and whether the mix has shifted.
- Competitor exposure. For every prompt, you see which competitors appear, with what frequency, and whose share is growing.
- Engine-by-engine breakdown. ChatGPT, Perplexity, Gemini, Google AI Mode, Copilot, Claude — each is tracked independently because, as Forrester argues elsewhere, guidance does not port between them.
The goal is not a single visibility score. The goal is a defensible answer to "what is happening, where is it happening, and what would change it." That answer is what survives a leadership conversation and a quarterly review.
Data for the decision, evidence for the conversation
Forrester closes his essay with a phrase worth stealing: data for decisions, evidence for conversations. Decisions need internal data — the spreadsheet, the time-series, the spike that triggers the investigation. Conversations need evidence — the screenshot, the cited snippet, the competitor's name surfacing in your prompt, the model's exact words.
Most AI visibility tools give you one without the other. A dashboard with no quotable evidence makes for a weak client call. A handful of screenshots with no underlying time-series makes for a weak board update. The pairing is the product.
If your team is still defending AI visibility budgets with anecdotes — "I saw ChatGPT recommend the competitor last week" — you do not have a content problem. You have a measurement infrastructure problem. The fix is not more content. The fix is the right evidence on the table when the conversation starts.
That is the work. Open the data, make it readable, give operators what they need to act. The machine holding the data changed. The discipline did not.
