Assistant · OpenAI

ChatGPT SEO for SaaS

ChatGPT SEO starts here. The largest assistant by a distance, and the one whose citation behaviour changes most between a model answering from memory and a model running a search.

51.3%of US assistant usage

Source: First Page Sage, Top Generative AI Chatbots by Market Share. July 2026, US monthly-active-user estimates.

A marketing lead reviewing results at a standing deskAssistant · OpenAI
Retrieval mechanism

Where a ChatGPT answer comes from

ChatGPT answers a question about your category in one of two modes, and they have almost nothing in common. In the first, the model answers from what it absorbed during training. There is no live fetch, no citation, and no way to influence the answer this quarter. In the second, it runs a search, retrieves pages, and builds an answer from what it just read, with links attached.

Almost everything that can be done for a brand in ChatGPT is work on the second mode, which is what ChatGPT search optimisation, still sometimes sold as SearchGPT optimisation, actually means.
The first mode is a long game measured in training cycles: what shapes it is how widely and how consistently your brand is described across the open web, not anything you can change on your own site this month.
The practical consequence is that the two modes fail differently.
If ChatGPT recommends a competitor without citing anything, you have a training-data and entity problem. If it runs a search and cites four sources that are not you, you have a retrieval problem, and retrieval problems are the ones that respond to work inside a quarter.
A growth lead planning the next quarter
Crawler policy

Which crawler does what, and what blocking it costs you

  • Training crawlers and retrieval crawlers are different things
  • Blocking a retrieval crawler removes you from the answers
  • Blocking a training crawler does not
RetrievalOAI-SearchBot
Builds the index behind OpenAI search, the results ChatGPT cites. Blocking it removes you from ChatGPT citations. This is the one that matters and the one most often blocked by accident.
User-triggered fetchChatGPT-User
Fetches a specific page when a user or a ChatGPT action follows a link. Blocking it breaks the experience for someone who has already found you.
TrainingGPTBot
Gathers training data. Blocking it is a legitimate commercial decision and does not remove you from search citations. It is a different question from the two above, and should be decided separately.

Our own policy is published at surge45.com/robots.txt, with a line per agent and the reasoning in the file.

A team in a meeting room with laptops open
Citation behaviour

How ChatGPT chooses what to cite

When ChatGPT searches, it retrieves a small set of pages and writes an answer from them, attaching links to the sources it drew on. The set is small: a handful of sources, not a page of ten blue links. The gap between being retrieved and being cited is narrow, but the gap between the top few results and everything else is enormous.

The sources it retrieves skew towards user-generated content more than the Google-side assistants do.
Community discussion and review platforms carry disproportionate weight for software-recommendation queries, which is why an on-site-only programme underperforms here even when the on-site work is good.
Answers are also conversational and multi-turn.
A buyer rarely asks one question. They ask for options, then narrow by price, then by integration, then by company size. Being cited on the first turn and absent by the fourth is a common and invisible failure, and it is only detectable if you test the whole sequence rather than a single prompt.
A leadership team reviewing numbers in a boardroom
What moves it

Observed to change ChatGPT visibility

01Presence on the sources ChatGPT retrieves
Reddit threads, G2 and Capterra listings, and comparison content on publisher domains. These are not your properties, which is exactly why most competitors are not working on them.
02Unambiguous entity definition
A model has to be able to say what you are in one clause. Categories that are described differently on your homepage, your G2 listing and your LinkedIn page produce answers that hedge or substitute a competitor with a clearer description.
03Answer-shaped content that survives extraction
Content that answers the question in the first two sentences of a section, with the qualifying detail after. A model quoting you should be able to lift one paragraph and have it stand alone.
04Freshness on comparison and pricing pages
Comparison queries retrieve heavily and a stale page loses to a current one. This is the least glamorous work on the list and reliably the most effective.
A team looking at results on a monitor together
What does not

Sold for ChatGPT, and not worth buying

✕ Blocking GPTBot to force citations
Training and retrieval are separate crawlers. Blocking the training crawler does not change search citations, and blocking the search crawler removes them entirely. This is the most common self-inflicted wound we find.
✕ Keyword density and volume plays
There is no ranking function to game. Retrieval selects passages that answer the question, then the model writes prose. More pages saying the same thing dilutes rather than compounds.
A team in a meeting room with laptops open
Working together

What we do on ChatGPT

  1. /01
    Build the real prompt set: the questions your buyers actually ask, including the follow-ups, not the head terms.
  2. /02
    Run them across the sequence and record which sources are cited at each turn.
  3. /03
    Separate the training-mode failures from the retrieval-mode failures, because they need different work on different timescales.
  4. /04
    Audit the OpenAI crawler policy in robots.txt, and fix the accidental blocks first.
  5. /05
    Prioritise the off-site sources that are being cited instead of you, then the on-site passages that would survive extraction.

ChatGPT questions we get asked

Can you guarantee ChatGPT will recommend us?

No, and nobody can. Model outputs are not deterministic and they change between model versions without notice. What can be done is to measure the citation rate for a fixed prompt set, work on the sources that are being cited instead of you, and re-measure. If someone offers you a guarantee here, that is the tell.

Should we block GPTBot?

It is a commercial question about training data, not a visibility question. Blocking GPTBot does not remove you from ChatGPT search citations. Blocking OAI-SearchBot does. Decide the two separately, and be certain which one your robots.txt currently blocks.

How long before anything changes?

Retrieval-mode visibility can shift within weeks, because it depends on pages that exist now. Training-mode behaviour moves on the vendor's release cycle and is measured in model generations. Any timeline that does not distinguish between the two is guessing.

Who can optimise a SaaS platform for ChatGPT?

Anyone willing to work on the search mode and the sources it retrieves, not just your own site. A ChatGPT optimisation agency should show you the prompt set, the sources cited at each turn and the OpenAI crawler settings it checked, then re-measure against the same set. Surge45 does this as part of GEO, run with your team every week and reported per assistant.

Ready when you are

Ready to build your GTM engine?

Book a discovery call with a GTM engineer. We'll identify which tier fits your stage and which part of the engine to build first.

A smiling strategist standing in front of a whiteboard of notes