Ten blue links became one answer.
Only a few sources get named in it.

The on-site work overlaps heavily. The retrieval surfaces do not. ChatGPT search draws on a narrower candidate set and cites fewer sources per answer, and observed citation patterns skew toward established publishers and recognisable brand coverage. Perplexity retrieves more aggressively at query time, names more sources per answer, and surfaces community discussion, documentation and Q&A threads far more often.

Same site work in both cases, different off-site emphasis, and separate measurement, because a gain on one platform does not imply a gain on the other. We track them as two lists rather than one score.

Yes, and more than the framing of AI SEO as a separate discipline suggests. Nothing gets cited that could not be retrieved, and retrieval still runs on crawlable, server-rendered, indexed pages. Most AI crawlers do not execute JavaScript the way Google's rendering service does, so content that appears only after hydration is often invisible to them.

The technical layer is a prerequisite rather than an alternative. What is genuinely different is the emphasis: entity clarity, passage-level self-containment and off-site corroboration carry more weight here than they do for a conventional ranking.

Entity SEO means making the identification of your business unambiguous. One consistent name, one canonical description, and the same founders, locations, services and credentials stated the same way everywhere they appear, with explicit sameAs links from your schema to the records that already describe you.

Google's Knowledge Graph is one such structured record, and it feeds Google's own AI surfaces directly. ChatGPT and Perplexity do not query it, but they are shaped by many of the same underlying sources that populate it, which is why Wikidata entries, company registries, review platforms and consistent press coverage keep turning up as leverage. Where two companies share a name, that ambiguity is usually the whole problem.

Not directly. A language model reads the text of a page rather than its JSON-LD. What schema does is remove ambiguity for the systems upstream of the model: it states plainly what a page is, who wrote it, what an entity is called, what a product costs and where a business operates, which improves how cleanly a page is parsed, categorised and pulled into a candidate set. It also drives the Google surfaces where structured records are read directly.

We treat it as accuracy infrastructure rather than a citation lever, and we keep it consistent with the visible text, because markup that contradicts the page is worse than no markup at all.

With a fixed prompt set and repeated sampling. These systems are non-deterministic, so one query proves very little: the same prompt run five times can cite five different sets of sources. We build a list of prompts matching how your buyers actually ask, run each of them on a schedule in both ChatGPT and Perplexity, and record how often you appear, whether you are named or merely paraphrased, which URL was cited, and who else is in the answer.

The figure that matters is the change in that rate over time, not any single screenshot. If you are assessing agencies for this work, ask for the prompt list and the baseline. Anyone doing it seriously has both, and can show which content changes moved which answers.

It depends which crawler, and the distinction is worth understanding before the decision is made. Different user agents do different jobs. Some collect training data, some build the search index an assistant retrieves from at answer time, and some fetch a page live because a user asked about it.

Blocking the training crawler while allowing the search crawler is a coherent position. Blocking the search crawler removes you from the citations you are trying to win. Most sites we audit have never made this choice deliberately, they inherited a robots.txt file from a template or a plugin.

It is a proposal for a plain-text file that summarises a site for language models, and the major assistants have not committed to reading it. We will add one on request, since the cost is close to zero, but we do not present it as a lever and we would not build a strategy around it.

Separating the parts of AI search work that demonstrably move citations from the parts that are still speculative is a large share of what an audit is for.

Faster than a Google ranking in some cases and slower in others. Perplexity retrieves at query time, so a corrected, well-structured page can begin appearing within weeks of being crawled. ChatGPT's search index refreshes on its own schedule and its citation set is more conservative, so movement there tends to track off-site coverage and takes longer.

Entity and authority work is the slow part, measured in months, because it depends on third-party sources updating their own records. We record the baseline in week one so the change is visible rather than argued about.