Original insights by Dan Petrovic
30-second rundown
Key learnings:
- AI answers depend on search indexes: If your page is absent from the index an AI system consults, you may never enter its consideration pool.
- Selection happens after retrieval: Topical centrality, engagement, popularity and persuasive on-page evidence influence which sources survive.
- Measure entities rather than imaginary prompt volume: Track brand mentions, linked mentions and citation share across a representative set of questions.
This week: Test five commercially important questions, record the brands and sources that appear, then improve one missing page or approach one influential outside source. Finish with a simple table showing whether your brand was absent, mentioned, linked or cited.
Coming 3 months: Use an automated AI Search monitoring workflow to generate related questions from your priority products and customer needs. Repeat them daily, track entity mention and citation share, and assign retrieval or on-page improvements when visibility changes.
Dan Petrovic opened the machinery of AI Search and found old search indexes rattling inside. The brochures promised a new oracle. Behind the curtain sat Bing, Google and Brave, feeding evidence into a mechanical judge with no patience for pages it could not find.
Ranking is only the first interrogation. After retrieval come grounding, mentions, citations and the final decision about which brand deserves to survive the answer.
1. The Machine Still Shops at Search Engines
Large language models do not simply roam the open web when answering a question. They retrieve candidates through existing search indexes, with ChatGPT leaning on Bing, Gemini using Google and Claude appearing to draw from Brave. If you are absent from the relevant index, you may never enter the consideration pool. The new oracle still buys its evidence from the old warehouse.
2. You Can Borrow a Source or Build One
Petrovic gives brands two routes into an AI answer. Earn inclusion on outside pages the system already trusts, or build your own page for the missing question. Start by inspecting the sources that already appear. The machine has left its guest list on the pavement, so there is no prize for pretending not to see it.
3. LLMs.txt Is Mostly a Ceremonial Shovel
A special file cannot rescue pages that search systems cannot discover, rank or trust. Petrovic sees little practical value in llms.txt because retrieval still depends on established indexes. Fix indexing and relevance before decorating the grave.
4. Three Old Ranking Forces Survived
Across search systems, Petrovic keeps returning to topical centrality, engagement and popularity. A site should be clearly associated with the subject, satisfy people when they arrive and attract genuine attention. AI Search did not kill authority signals. It merely dragged them into a hotter, stranger courtroom.
5. Google Is the Base Camp, Not the Entire Mountain
The engines use different indexes, but Petrovic does not recommend splitting the company into three frantic optimization cults. Strong SEO creates the common foundation, after which important gaps can be checked in Bing or Brave. Build broadly strong pages, then investigate specific absences.
6. AEO Begins Where Ordinary Ranking Stops
Retrieval gets a source through the door. The model must still ground its answer in that evidence, mention the brand and decide whether to attach a citation. The richest outcome is a linked brand mention in the LLM answer. A citation without the brand mention can leave your page serving drinks at a competitor’s victory party.
7. Training Data Is the Slow, Foggy Road
Brands may influence what models remember through widespread, consistent information, but training cycles are opaque and difficult to measure. Web retrieval offers a faster and more observable route. Optimize the evidence models can fetch today before betting the farm on what a future model might remember.
8. Compression Can Smell the Content Factory
Petrovic explores Gzip compression as a rough way to separate information-rich writing from repetitive spam. Formulaic pages compress unusually well because the same language keeps returning in a cheap suit. Repetition leaves a mechanical fingerprint. Scaling near-duplicate copy may therefore advertise the very emptiness it was designed to conceal.
9. Make the Page Difficult to Resist
Petrovic uses an agentic loop in which a model reviews a page, explains why it would reject it and guides another revision. The process continues until the page answers the question with clearer evidence and fewer objections. Use AI as the hostile selector, not merely the copywriter.
10. There Is No Honest Prompt Search Volume
Prompt wording is effectively infinite, so a precise monthly search volume for one prompt is theatre. Petrovic works outward from entities and fan-out questions, then measures mention share and citation share across a representative set. Track the market of questions, not one sacred sentence.
The machine does not need another thousand pages of perfumed filler. It needs discoverable evidence, a credible reason to choose you and enough repeated testing to separate a pattern from a hallucination. Build that, measure the entity and leave the ceremonial AEO shovels rusting beside the highway.
