Most people still start with a search box. A growing share now type directly into ChatGPT or Perplexity and read the synthesized answer the model writes back. If your site is not named in that answer, that reader never finds you.
The good news is that the work is concrete. Both engines have patterns you can optimize for, and most of the gains come from content decisions you can act on this week.
How ChatGPT and Perplexity retrieve sources
Understanding the retrieval pipeline matters before you try to influence it.
Perplexity is a live-retrieval-first engine. When a user asks a question, it runs a real-time web search, pulls a set of candidate pages, reads them, and writes a synthesized answer that cites each source it drew from. The citations are visible and prominent. Being cited in Perplexity looks a lot like a footnoted source in a research report.
ChatGPT blends two inputs. Its training data, baked in months before any query, shapes what the model already knows about a topic. When web search is enabled, which it is in paid tiers and many free experiences, it also runs live retrieval similar to Perplexity’s. The weighting between the two shifts by topic and by how current the information is. For a fast-moving niche, live retrieval likely dominates; for an established topic, training data plays a larger role.
The practical implication: content that earned authority before a model’s training cutoff can pay dividends through training data alone, while fresh, crawlable pages feed the live-search pipeline. Both matter. Neglecting either closes half the aperture.
If the model cannot fetch your page, or cannot read it quickly, you are not in the candidate set. Getting cited downstream starts with accessibility upstream.
What both engines reward
Despite the architectural differences, the two engines converge on what they value in a source.
Quotable, self-contained claims. A model assembles an answer from fragments. A sentence that makes sense without the surrounding paragraph travels well into a synthesized reply. Buried arguments rarely get lifted; obvious answers, stated plainly in the first line of a section, often do.
Entity clarity. Answer engines reason over entities: named organizations, products, and people, and how they connect to topics. If your site does not clearly state who you are and what you cover, the model has to infer it. Inferences get omitted in favor of clearer sources.
Corroboration. A claim that matches what other credible sources say is more likely to appear in a generated answer than a lone-voice assertion. Your differentiated positions should be backed by evidence the model can see; your statements on established facts should align with the record.
Crawlable, fast pages. Perplexity’s live retrieval and ChatGPT’s search mode both need to fetch and read your pages under real-world conditions. A slow server, aggressive caching rules, or a robots.txt that accidentally blocks crawlers removes you from the running. Auditing what AI crawlers can access is a quick check worth doing before anything else.
Tactics, ordered by return
The following is not a universal launch sequence. Prioritize by what your site is actually missing.
Lead with the answer
The single highest-return change is putting the answer at the top of the section that covers a question, rather than building toward a conclusion at the end. State the answer in one or two sentences. Then expand. The model reads the first passage in a section and decides whether to use it; readers who want depth scroll down.
Headings phrased as questions reinforce this. They tell the retrieval system what the section covers and match the query patterns users actually type.
Remove extraction friction
Keep paragraphs short. Write definitions that stand alone. Avoid prose where meaning depends on the prior three paragraphs for context. Lists work well for process steps and comparisons, because each item is an independent unit the model can lift without carrying the rest of the page with it.
Be consistent about your entity
Pick names and stick to them. If you call your service “AI visibility consulting” on one page and “AI search optimization” on another, you split your signal across two entities the model may not connect. Use the same term for the same thing everywhere, and make at least one page state explicitly what your organization does and which topics it covers. Schema markup reinforces this at a machine-readable level, but the prose is the foundation.
Publish an llms.txt file
llms.txt is a plain-text file at your domain root that tells AI systems which pages are most worth reading and how to interpret your site. The format is still settling, and not every engine consumes it yet. It is cheap to ship and signals exactly the kind of legibility these systems favor. llms.txt explained covers the format, the limits, and what to include.
Audit your robots.txt
Check that major AI crawlers are not blocked unintentionally. User-agent rules written to exclude aggressive scrapers a few years ago may also exclude GPTBot, PerplexityBot, or ClaudeBot today. A page blocked to a crawler cannot be cited.
Where ChatGPT and Perplexity diverge in practice
The tactics above apply to both engines. A few differences are worth tracking once you are past the basics.
Perplexity shows its sources by default, with numbered links in the answer. Being cited there is legible: the user can see it and click it. Your page title and the first sentence of the cited passage matter for how your result appears in that list.
ChatGPT’s behavior varies by context. In a conversation without search enabled, it draws on training data and may describe your brand based on what it learned before its cutoff, accurately or not. In search-enabled mode, it behaves more like Perplexity. The audit discipline of checking what models say about you, then publishing clear, citable content that corrects the gaps, matters more for ChatGPT than for Perplexity, where every answer is freshly retrieved.
Measuring whether it is working
There is no rank position to track. Run citation audits instead.
Keep a list of the questions your buyers ask ChatGPT and Perplexity. Run each one monthly and record whether your site appears, whether you are named, and which competitors show up instead. Track the trend across that question set; a rising share of mentions is the signal that the work is landing.
Watch referral traffic from AI engines in your analytics. Volume is still modest but growing, and sessions from AI engines tend to show high intent. A drop in AI referral traffic while organic holds steady can indicate a crawl or citation issue worth investigating before it widens.
For a fuller measurement framework, measuring AI search traffic covers the specific signals to watch and how to read the numbers without over-interpreting early data.
How Strynal approaches AI visibility
We start every engagement by asking what your buyers are actually searching for in ChatGPT and Perplexity today, then checking whether your site appears in those answers. In most cases it does not, for fixable reasons: content buried under abstract headings, an entity the model cannot reliably connect to a topic, or crawl rules that exclude the wrong bots.
The work spans content strategy, technical accessibility, and structured data. It is the same seam a focused studio is built to own, because the team that diagnoses the problem ships the fix.
The broader context is in what generative engine optimization actually is, if you want to understand how AI answer engines select and cite sources in the first place. For the engagement itself, our AI visibility practice is the place to start. We can show you who the models cite today in your space, and what it takes to be the answer.