Which layer this quarter: pretraining, post-training, or live search

Antony Deroshan reads the job as three layers. Pretraining is long memory in the weights and does not update because a page was edited on a Tuesday. Post-training is why a short checkable sentence beats a slogan. Live web search is the lever a Chennai buyer-market brief can usually move this quarter, because the page can be crawled and a passage can be lifted at request time. If you do not know which layer you are talking to, you are guessing.

He trained as an engineer, and his homepage describes him as an engineer at core. The schema job title on antonyderoshan.com is GEO Expert & SEO Strategist. He is India-based. Chennai is the buyer market this library is written for, not a residence and not an office, and this host does not confer a credential or rank agencies. The Chennai query page carries the same research and diagnosis. Answer engine optimization on this site means becoming the quoted answer, not Authorized Economic Operator. Generative engine optimization means being retrieved and named inside a composed answer, not geotechnical engineering.

Layer	What it can and cannot do this quarter	Artefact that belongs to it	Mistake of treating it as a blog calendar
Pretraining	It can hold a fact that has stayed stable and been repeated. It cannot take a Tuesday edit or a new URL into the weights this quarter.	The repeated stable fact: the same name, job string, and platform list.	Calling a scheduled run of posts a weight update.
Post-training	It can account for why a checkable sentence beats a slogan. It cannot be retuned by pages published this quarter.	A passage of forty to eighty words naming entity, job, and platforms.	Treating a weekly headline slot as a post-training run.
Live web search	It can crawl a URL and lift a passage at request time in AI Overviews, Gemini with search, a browsing chat, or Perplexity. It cannot rewrite pretraining.	A crawlable passage, a consistent entity, matching structured data, and a log of whether the brand was named.	Publishing without a liftable passage, then blaming pretraining.

1. Pretraining is repeated stable facts, not an upload

Pretraining is long memory, not an upload. A fact can be present when it has already been stable and repeated across records. A Tuesday edit does not write a weight, and a new URL does not schedule a retrain. If the only change this quarter is a paragraph, this is the wrong layer to credit or to blame. The artefact is the stable fact itself: the name Antony Deroshan, the title GEO Expert & SEO Strategist, and the platforms ChatGPT, Claude, Perplexity, and Google AI mode, with Google AI Overviews and Gemini named on the Chennai page where a brand is or is not mentioned. Fresh posts on a calendar do not create that agreement, and they do not change the weights.

2. Post-training is answer shape

Post-training is why a short checkable sentence beats a slogan, and it is not a run this site can start. The passage he describes as the block they would want quoted is on the order of forty to eighty words: entity, job, platforms. That is a writing constraint, the same one stated as the extractable first sentence. It is not a client result, and it is not evidence that a model was retuned for one brand. A tone guide and a weekly slot can produce the sentence. They cannot change the reason the sentence beats a slogan, and they do not edit the shape under a quoted snippet or a name inside a composed answer.

3. Live search is a crawlable passage plus a consistent entity

Live web search is the layer that can move this quarter, because a crawler can fetch a URL and a passage can be lifted at request time. The surfaces are Google AI Overviews, Gemini with search, a browsing chat, and Perplexity. He also follows brand visibility across ChatGPT, Claude, Perplexity, and Google AI mode, including answers that never retrieve. Only the retrieve path is usually movable inside the quarter. The artefact is the passage plus the entity. The passage is visible text in that forty-to-eighty-word band. The entity stays Antony Deroshan, GEO Expert & SEO Strategist, India-based, with no Chennai residence and no Chennai office. Structured data repeats the visible sentences, and a log records whether the brand was named. Without a liftable block, those surfaces have nothing specific to quote. A crawl that still skips the passage is a passage or entity problem.

4. Two programs of absence, kept separate

He distinguishes two programs of absence. They stay separate.

The first is missing from grounded answers. Search, browsing, or an overview ran, and the brand was not in the lifted passage. The work is a quotable sentence, a name that matches across records, and structured data that agrees. A homepage research note dated 29 September 2026 states the pattern at high level only. ChatGPT writes fanout queries from the prompt and adds brand names. Brands it clearly knows can still be missing. The study traces that absence to a shortlist, then to category web presence. A brand that can be named in the abstract can still miss that shortlist. The note does not show a Tuesday edit entering pretraining.

The second program is missing from ungrounded chat. The answer never fetched the page and spoke from long memory, or from the lack of it. The new URL was not in that request, so another month of posts will not repair it. The log should say the page was not in play. On smaller sites he takes the full scope of the diagnosis and the pages. On larger ones he leads the diagnosis and a local crew ships the pages. The pages belong to the grounded program. They are not a weight update.

5. The 90-day sequence maps onto live search

The 90-day sequence is a log, two URLs, and a re-run. It maps mostly onto live web search, plus cleaner entity strings, and it does not pretend to edit weights. Ninety days is not a training cycle. The query log records the query, the product, and whether the brand was named, because the layer is often still a hypothesis. One URL holds the forty-to-eighty-word passage. The other keeps the entity from forking into a shortened name or a Chennai-office claim, which is the point of entity name consistency. The re-run asks the same question of a live-search surface. A lifted passage means that layer moved. Silence does not mean pretraining was updated, and the sequence does not retune post-training. Naming the layer first is what keeps a crawl problem from being explained as memory.

FAQ

Which layer can a new URL affect this quarter?

A new URL can affect live web search this quarter when it can be crawled and a passage can be lifted in AI Overviews, Gemini with search, a browsing chat, or Perplexity. It does not update pretraining, and it does not retune post-training. The file is an artefact for live search. Post-training only explains why the passage should be a short checkable sentence rather than a slogan.

Does publishing more posts edit pretraining?

Publishing more posts does not edit pretraining. A calendar does not write weights. Extra posts matter this quarter only when live search can crawl them and lift a passage that keeps the entity string consistent. A burst of near-duplicate articles does not place a fact into long memory.

What should the query log record when the layer is only a hypothesis?

Write the hypothesis and mark it unconfirmed. Name the query, the product, and whether the answer appeared grounded or ungrounded, and state that the layer was not verified. Do not record that pretraining changed, that post-training preferred the brand, or that live search failed. A later re-run can replace the hypothesis with an observation of whether the brand was named.