Ask a question in ChatGPT, Perplexity, Gemini, or Copilot, and the answer appears in seconds. It feels effortless. But under the hood, there’s no magic. There’s a fight happening.
This is the part of the pipeline where your content is in a knife fight with every other candidate. Every passage in the index wants to be the one the model selects.
For SEOs, this is a new battleground. Traditional SEO was about ranking on a page of results. Now, the contest happens inside an answer selection system. And if you want visibility, you need to understand how that system works.
Image Credit: Duane ForresterThe Answer Selection Stage
This isn’t crawling, indexing, or embedding in a vector database. That part is done before the query ever happens. Answer selection kicks in after a user asks a question. The system already has content chunked, embedded, and stored. What it needs to do is find candidate passages, score them, and decide which ones to pass into the model for generation.
Every modern AI search pipeline uses the same three stages (across four steps): retrieval, re-ranking, and clarity checks. Each stage matters. Each carries weight. And while every platform has its own recipe (the weighting assigned at each step/stage), the research gives us enough visibility to sketch a realistic starting point. To basically build our own model to at least partially replicate what’s going on.
The Builder’s Baseline
If you were building your own LLM-based search system, you’d have to tell it how much each stage counts. That means assigning normalized weights that sum to one.
A defensible, research-informed starting stack might look like this:
- Lexical retrieval (keywords, BM25): 0.4.
- Semantic retrieval (embeddings, meaning): 0.4.
- Re-ranking (cross-encoder scoring): 0.15.
- Clarity and structural boosts: 0.05.
Every major AI system has its own proprietary blend, but they’re all essentially brewing from the same core ingredients. What I’m showing you here is the average starting point for an enterprise search system, not exactly what ChatGPT, Perplexity, Claude, Copilot, or Gemini operate with. We’ll never know those weights.
Hybrid defaults across the industry back this up. Weaviate’s hybrid search alpha parameter defaults to 0.5, an equal balance between keyword matching and embeddings. Pinecone teaches the same default in its hybrid overview.
Re-ranking gets 0.15 because it only applies to the short list. Yet its impact is proven: “Passage Re-Ranking with BERT” showed major accuracy gains when BERT was layered on BM25 retrieval.
Clarity gets 0.05. It’s small, but real. A passage that leads with the answer, is dense with facts, and can be lifted whole, is more likely to win. That matches the findings from my own piece on semantic overlap vs. density.
At first glance, this might sound like “just SEO with different math.” It isn’t. Traditional SEO has always been guesswork inside a black box. We never really had access to the algorithms in a format that was close to their production versions. With LLM systems, we finally have something search never really gave us: access to all the research they’re built on. The dense retrieval papers, the hybrid fusion methods, the re-ranking models, they’re all public. That doesn’t mean we know exactly how ChatGPT or Gemini dials their knobs, or tunes their weights, but it does mean we can sketch a model of how they likely work much more easily.
From Weights To Visibility
So, what does this mean if you’re not building the machine but competing inside it?
Overlap gets you into the room, density makes you credible, lexical keeps you from being filtered out, and clarity makes you the winner.
That’s the logic of the answer selection stack.
Lexical retrieval is still 40% of the fight. If your content doesn’t contain the words people actually use, you don’t even enter the pool.
Semantic retrieval is another 40%. This is where embeddings capture meaning. A paragraph that ties related concepts together maps better than one that is thin and isolated. This is how your content gets picked up when users phrase queries in ways you didn’t anticipate.
Re-ranking is 15%. It’s where clarity and structure matter most. Passages that look like direct answers rise. Passages that bury the conclusion drop.
Clarity and structure are the tie-breaker. 5% might not sound like much, but in close fights, it decides who wins.
Two Examples
Zapier’s Help Content
Zapier’s documentation is famously clean and answer-first. A query like “How to connect Google Sheets to Slack” returns a ChatGPT answer that begins with the exact steps outlined because the content from Zapier provides the exact data needed. When you click through a ChatGPT resource link, the page you land on is not a blog post; it’s probably not even a help article. It’s the actual page that lets you accomplish the task you asked for.
- Lexical? Strong. The words “Google Sheets” and “Slack” are right there.
- Semantic? Strong. The passage clusters related terms like “integration,” “workflow,” and “trigger.”
- Re-ranking? Strong. The steps lead with the answer.
- Clarity? Very strong. Scannable, answer-first formatting.
In a 0.4 / 0.4 / 0.15 / 0.05 system, Zapier’s chunk scores across all dials. This is why their content often shows up in AI answers.
A Marketing Blog Post
Contrast that with a typical long marketing blog post about “team productivity hacks.” The post mentions Slack, Google Sheets, and integrations, but only after 700 words of story.
- Lexical? Present, but buried.
- Semantic? Decent, but scattered.
- Re-ranking? Weak. The answer to “How do I connect Sheets to Slack?” is hidden in a paragraph halfway down.
- Clarity? Weak. No liftable answer-first chunk.
Even though the content technically covers the topic, it struggles in this weighting model. The Zapier passage wins because it aligns with how the answer selection layer actually works.
Traditional search still guides the user to read, evaluate, and decide if the page they land on answers their need. AI answers are different. They don’t ask you to parse results. They map your intent directly to the task or answer and move you straight into “get it done” mode. You ask, “How to connect Google Sheets to Slack,” and you end up with a list of steps or a link to the page where the work is completed. You don’t really get a blog post explaining how someone did this during their lunch break, and it only took five minutes.
Volatility Across Platforms
There’s another major difference from traditional SEO. Search engines, despite algorithm changes, converged over time. Ask Google and Bing the same question, and you’ll often see similar results.
LLM platforms don’t converge, or at least, aren’t so far. Ask the same question in Perplexity, Gemini, and ChatGPT, and you’ll often get three different answers. That volatility reflects how each system weights its dials. Gemini may emphasize citations. Perplexity may reward breadth of retrieval. ChatGPT may compress aggressively for conversational style. And we have data that shows that between a traditional engine, and an LLM-powered answer platform, there is a wide gulf between answers. Brightedge’s data (62% disagreement on brand recommendations) and ProFound’s data (…AI modules and answer engines differ dramatically from search engines, with just 8 – 12% overlap in results) showcase this clearly.
For SEOs, this means optimization isn’t one-size-fits-all anymore. Your content might perform well in one system and poorly in another. That fragmentation is new, and you’ll need to find ways to address it as consumer behavior around using these platforms for answers shifts.
Why This Matters
In the old model, hundreds of ranking factors blurred together into a consensus “best effort.” In the new model, it’s like you’re dealing with four big dials, and every platform tunes them differently. In fairness, the complexity behind those dials is still pretty vast.
Ignore lexical overlap, and you lose part of that 40% of the vote. Write semantically thin content, and you can lose another 40. Ramble or bury your answer, and you won’t win re-ranking. Pad with fluff and you miss the clarity boost.
The knife fight doesn’t happen on a SERP anymore. It happens inside the answer selection pipeline. And it’s highly unlikely those dials are static. You can bet they move in relation to many other factors, including each other’s relative positioning.
The Next Layer: Verification
Today, answer selection is the last gate before generation. But the next stage is already in view: verification.
Research shows how models can critique themselves and raise factuality. Self-RAG demonstrates retrieval, generation, and critique loops. SelfCheckGPT runs consistency checks across multiple generations. OpenAI is reported to be building a Universal Verifier for GPT-5. And, I wrote about this whole topic in a recent Substack article.
When verification layers mature, retrievability will only get you into the room. Verification will decide if you stay there.
Closing
This really isn’t regular SEO in disguise. It’s a shift. We can now more clearly see the gears turning because more of the research is public. We also see volatility because each platform spins those gears differently.
For SEOs, I think the takeaway is clear. Keep lexical overlap strong. Build semantic density into clusters. Lead with the answer. Make passages concise and liftable. And I do understand how much that sounds like traditional SEO guidance. I also understand how the platforms using the information differ so much from regular search engines. Those differences matter.
This is how you survive the knife fight inside AI. And soon, how you pass the verifier’s test once you’re there.
More Resources:
- AI Search Changes Everything – Is Your Organization Built To Compete?
- Why Generative AI Isn’t Killing SEO – It’s Creating New Opportunities
- SEO In The Age Of AI
This post was originally published on Duane Forrester Decodes.
Featured Image: tete_escape/Shutterstock
