Position 15 in Google pays nothing. Ethan Smith, who runs the agency Graphite, on the Ahrefs Podcast: “All that matters in search is, are you in the top three? I don’t care about any other rank really, because you don’t get any clicks.” Then the reversal, in the same breath: “Position 10 and position 15 in citations can matter a lot.” Same slot number, opposite value.
The difference is geometry. A Google results page is a ranked list you scan and abandon. An AI answer is assembled from a retrieval set, and the model reads the whole set before it writes a word. What “position” means depends on which of those two objects you occupy.
How many citations does a RAG answer read?
About 20. Smith, describing the retrieval step in the same episode: “it’s not just looking at three citations. It’s looking at 20. It’s looking at many citations.” Every page in that set feeds the answer. Slot 15 is inside the room. In Google, slot 15 lives on page two, a place people cite in jokes and do not visit.
A second channel sits on top of retrieval. The core model carries co-occurrence patterns from training: which brands appear next to which category terms across the web. Smith again: “even if that’s not in the citation set, the core model sees that and the more, the better.” Presence compounds twice. Once in the retrieval set, once in the weights.
What does exclusion cost?
Zero is the number. An AI answer lists 2 to 4 options and stops; Google offers infinite pages and lets you fall down them. Dharmesh Shah, HubSpot’s co-founder, on the stakes: “if you’re not in the citations, you’re literally at zero.” Google demotes you in increments. An answer omits you in full.
That makes AI visibility binary in a way rank tracking never was. You are one of the few names in the paragraph, or you are a company the buyer never hears about. There is no position 8 consolation prize inside a sentence.
How deep do engines read?
Deeper than any human. One shower-head purchase query pulled 65 web pages into ChatGPT’s context, per James Cadwallader of Profound on Sequoia’s podcast. A person reads the top 4 or 5 links. Agents traverse the tail: the comparison page parked at position 40, the help-center article no session recording has ever touched. Your analytics call those pages dead. The retrieval layer calls them sources.
Does your Google rank transfer?
About a third of it. ChatGPT’s citations overlap Google’s results at around 35% per Graphite’s data; Perplexity overlaps Google at around 70%. Two thirds of what ChatGPT cites in your category did not come from the Google results you spent a decade optimizing. Different index, different winners.
So the intervention list changes. Pages that never had a shot in Google can win citations: pricing pages, help-center articles, integrations pages, tail content answering questions with no search volume. Smith sees “a lot of small sites ranking in citations, especially for the long tail” (Surfer Academy). A page that lost its Google fight in 2022 might be load-bearing in an answer today. Nobody told it.
The tail cuts the other way too. In the Surfer session, Smith walked through a website-builder answer where Squarespace had five distinct URLs feeding a single response: a feature page, a checklist, a tools page, a help-center article, a how-to. His comment: “I don’t think Squarespace would have five URLs ever appearing in Google for a non-branded search like this, but it does work here.” Google caps a domain at one or two slots per page of results. A citation set has no such manners.
Expect the set itself to keep churning. Smith’s diagnosis of why odd sites show up in citations: “it’s not that hard to build a search engine that has a good relevance score. It’s very hard to sort… very hard to build the authority score.” The engines are years from a settled PageRank equivalent, so “the citations that appear will be very different over the coming years.” His advice is to go after all of them. Today’s citation list is a snapshot of an immature sorting algorithm, and the sorting algorithm does not yet share Google’s opinion of your site.
Most teams have never looked at their citation sets, which is the part you can fix this week. Our free scan asks 5 questions your buyers ask and shows every citation the engine returned, answer text included. The paid audit runs each question 7 times per engine, because a single run is a snapshot, never a rank, and reports your presence as a range. Positions 5 through 20 stopped being noise the day the reader became a machine that takes 20 sources. Google trained you to ignore them. Retrain.
Sources
- Ethan Smith on the Ahrefs Podcast
- Ethan Smith at Surfer Academy
- Dharmesh Shah on My First Million, Sep 2025
- James Cadwallader (Profound) on Sequoia’s podcast, Apr 2026