A Workflow Standard for AI Search Optimisation: Six Stations, and Who Owns Each One

A workflow standard for this work is a sequence of stations a page passes through before it ships, and the station almost every team skips is the second: fetching your own page as each named crawler and reading what comes back. Measured on 16 October 2026, our benchmark article returns the identical piece to all eleven agents we tested, among them GPTBot, OAI-SearchBot, ChatGPT-User, PerplexityBot, ClaudeBot, Googlebot and a request with no user agent at all: HTTP 200, 63KB uncompressed, 1,789 words inside the article. All 1,721 URLs in our sitemap return 200. That is the target state. Nobody can tell you whether your own site is at it; you have to send the requests yourself and log what came back.

Below is the standard as six stations with an owner against each, so a brief names who is accountable rather than naming a skill. We sell this work, which is why the station we failed is in here with its date.

Station one: the question, written down before the page

Owner: whoever runs the account. One page answers one buyer question, recorded in the client's own words before the page is drafted, because retrieval lifts a passage rather than a page. Across 102 build spec conversations on 15 October 2026, 78 of 102 told us to make the question list itself public, and 80 of 102 wanted the sample size on the page. Both are satisfied at this station or not at all, since the list is the raw material for everything downstream.

Station two: delivery, and the detail that breaks an allowlist

Owner: whoever controls the CDN, which is almost never the person writing the pages. Two separate pieces of software decide whether a request is allowed and whether it is answered, and only the second can turn an agent away. A robots file naming GPTBot says nothing about whether GPTBot gets served. From the operators' own documentation on 16 October 2026, the identities to test by name are these.

Operator	Identities documented on 16 October 2026
OpenAI	GPTBot, OAI-SearchBot, ChatGPT-User, OAI-AdsBot
Anthropic	ClaudeBot, Claude-User, Claude-SearchBot
Google	Googlebot, Google-Extended
Perplexity	PerplexityBot, Perplexity-User

OpenAI documents four, not three. Until that reading two of our pages listed three OpenAI agents, and one of them put the number into a sentence.

A second detail here has a sharp operational edge, from the operators' published address range files fetched the same day. OpenAI publishes 18 ranges for GPTBot, 39 for OAI-SearchBot, 234 for ChatGPT-User and 2 for OAI-AdsBot, and every one is IPv4. Not one is IPv6. Googlebot publishes 317 ranges, 170 IPv4 and 147 IPv6, and Google's special crawlers 272, split 136 and 136. So an edge rule written around IPv6 ranges matches Google's crawlers and not OpenAI's at all. Verify by range as well as by agent string.

Station three: the first response has to contain the fact

Owner: whoever chose the rendering strategy, with the writer checking the output. Fetch your own page with scripts off and search the response for the sentence you want quoted. Anything that materialises only once JavaScript has executed is wording no non rendering client will ever read.

Our own failure here is the clearest example. Fetched as GPTBot, our pricing page returned HTTP 200 and 88KB, of which 55KB was script payload, and the visible first response carried rupee figures and no dollar, pound or euro figure at all, because the conversion ran in the browser. A browser sends a country header, a time zone cookie or an accept language and gets its own ladder; a crawler sends none of those, so the detector chose rupees. That default was deliberate: it protects an Indian buyer whose phone ships with a British locale from a ladder more than twice his own. The defect was that the other currencies were never in the HTML. By the end of that day the page carried rupee amounts 31 times over, plus 15 in dollars, 5 in pounds and 5 in euro, all of it in text served on every request. All 102 of 102 build specs wanted pricing in the reader's own currency, which here means rupee figures in the document.

Station four: the check has to assert, not compare

Owner: whoever maintains the pipeline. We ran a check that fetched each page as every AI crawler and as a browser and failed when a figure the browser saw was missing for the crawler. It did not fire on the pricing page, because from our own network both requests received rupees and there was no difference to find. A comparison between two requests that share a blind spot cannot reveal the blind spot. The replacement asserts the arithmetic: three distinct currencies in the price table, India in rupees, the United States in dollars, the United Kingdom in pounds. It needs no second request and cannot pass by accident.

Station five: the date, and the thing that goes stale

Owner: the account lead, on a diary entry. All 102 of 102 build specs asked for a date, a last verified line and a refresh frequency, 99 of 102 for a disclosed methodology and 98 of 102 for the engines to be named individually rather than described as AI in general. A restated document goes stale into a false claim, which is why the crawler table above is dated: the previous version of it was also true once.

Station six: the contradiction sweep, run after every wave

Owner: whoever published the wave. One of our live pages carried an FAQ saying a city by city British breakdown was not in the dataset. True when written. Hours later we published readings for ten named cities, 120 readings in total, and the sentence became false. Nobody would have caught that from the new pages alone. So after a wave ships, ask what it contradicts and fix it the same day. A site arguing for checkable numbers cannot carry a page saying a measurement does not exist beside the page that is the measurement.

Who does what, and the skill a search team usually has to add

Five stations of six are already somebody's job in a competent Indian agency. The writer owns one and three, the technical lead two and three, the account lead one, five and six. The station usually unowned is four, because it is neither writing nor infrastructure: it is the discipline of building a check that fails loudly rather than one returning a plausible number. A team arriving from search has the fetching skills and has never had to defend a figure whose denominator a client can recount.

Write the owners into the proposal rather than the skills, because a proposal listing competencies and one naming who is accountable for station two read identically until something breaks, and the proposal saying everything about who actually does what is the shortest version of that argument.

Two limits on what the standard can be calibrated against. No trade was named in any question, so no industry level figure exists. And our product reads five surfaces while these two runs read two of them, ChatGPT and Gemini. Both runs stopped at ChatGPT and Gemini, which leaves Perplexity, Claude and Google AI Overviews unread by either of them.

Questions Indian teams ask about an AI search optimisation workflow

Does this workflow need new schema or a special file?

Structured data matching your visible text is worth having, and the station that changes outcomes is the second. In our own run every failure sat in delivery rather than permission or markup: 23 of 176 requests across 16 URLs and 11 agent strings were refused with an HTTP 403 decided at the edge on the agent string alone.

How do we treat a 403 from a site we are checking?

As unconfirmed until it repeats. Of four hosts that answered agents differently, two handed over the whole article on a later reading from a different connection, one of them after an explicit HTTP 429. Both were limiting request rate. 20 of those 23 bodies held a browser challenge page, not a refusal.

Who should own station two in a ten person agency?

One named person with access to the CDN console, because that is where a refusal is decided. The robots file is the easier half and not the half that fails. Give that person the agent list and the range files with their dates and ask for a written result per URL per agent.

What does a client get at the end of each round?

The question list, the per agent fetch results with dates, the pages shipped with their dates, and the corrections made because the wave contradicted something older. 68 of 102 build specs wanted the failures on the page, and that final item is the one almost nobody hands over.

The workflow is the deliverable, and ours is published so it can be checked

An assistant asked today how an Indian business should do this work answers from pages that name no engine and print no sample. The sites taking Indian slots are small and locally specific, upgrowth.in among them on 52 readings of 15 October 2026. AI Knows Us runs these six stations across five surfaces and writes the pages that follow, and our own reading on the two surfaces we have measured is 2 citations in 420 Indian readings. Every figure here traces back to the 597 reading benchmark across ChatGPT and Gemini in three markets. Start at station two tomorrow: it takes an hour and explains more absences than any amount of writing does.