Home / Guides

How to rank in ChatGPT: what actually moves a brand into an answer

There is no ranking table inside ChatGPT, so the honest version of this question is what makes the assistant name your brand when a buyer asks it for options. This guide walks through where an answer comes from, which moves plausibly change it, which are folklore, and how to measure the result without fooling yourself.

Seenby is not open yet, so anything on this page about what it does is what it is being built to do.

The short answer: you do not rank, you get named

ChatGPT returns a paragraph, not an ordered list. There is no index position to climb. No page two. No number to check. Your brand is either in the paragraph a buyer reads or it is not. So the question worth working on is narrower: what makes an assistant name you when someone asks for options in your category.

Presence has three parts, and they respond to different work. Whether you are named at all. Where you land when you are named, first as the default or fourth among also-rans. And how you are characterised. "The enterprise option, expensive" and "the one mid-market teams pick" are not the same mention. Most teams measure only the first part. Then they argue about the third in a meeting with no evidence.

Here is the honest boundary of what you can influence. You cannot edit the model. You cannot see the routing, the system prompt or the retrieval behind any single answer, and the providers do not publish it. You can influence what the public record says about your category. You can influence how clearly you are described as an entity. You can influence whether your own pages give a retrieval pass something concrete to quote. Everything below sits inside that boundary.

One instruction before the method. Take a baseline before you change anything. Answers differ between identical asks, and they move when the model behind them is updated. Without a before, every after is a story you tell yourself.

Where an answer about your category comes from

Three inputs produce the paragraph. Separating them tells you which work pays back in years and which pays back in weeks.

The model's trained memory

The model absorbed a large amount of public text before anyone asked it anything. It can name brands in your category with no live lookup at all. This is the slow layer. You cannot edit it. You cannot request a correction. The next training run is not scheduled around your campaign. What you can do is change what the public record says over time, which is the same patient work as reputation, on a longer fuse.

Retrieval at answer time

When the assistant searches before answering, it pulls a small set of pages and writes from them, usually showing the links it used. This is the fast layer, and it is where most of the practical work lives. If the pages that already answer "best tools for X" do not mention you, a retrieval pass will not invent you. If they do, you are in the material the answer is written from.

The prompt, and everything around it

The user's phrasing, the previous turns, their location and any saved instructions all shape the output. Two buyers asking what looks like the same question get different lists. This matters less for strategy than for measurement. If you change the question every time you check, you are measuring your own curiosity, not the market. Fix the wording, then watch it.

The surfaces you can actually work on

The third-party consensus about your category

Comparison posts, review directories, analyst round-ups, community threads and forum answers. These are both the training material of tomorrow and the retrieval pool of today. If the pages that already rank for your category question do not carry your name, the assistant has no reason to produce it.

Your own pages, written to be quoted

Plain sentences that state who the product is for, what it does, what it costs, what it does not do. A paragraph generator rewards text that can be lifted and paraphrased without interpretation. Marketing abstraction gives it nothing to hold on to.

Entity clarity

One consistent description of what your company is, which category it sits in, and which buyer it serves. Repeat it wherever you are described. Ambiguity gets you merged with a similarly named company or dropped from a category you belong in. This is unglamorous, and it is often the highest-yield fix.

Specific, checkable facts

Numbers, limits, supported platforms, named use cases. Facts survive summarisation better than adjectives. They are what an assistant reaches for when a buyer asks a narrow question. They also reduce the odds of an inaccurate description, because there is something concrete to describe.

A working method, in the order you need it

Write down the prompts that matter
Not keywords, questions. The way a buyer actually asks: best tool for this job, alternatives to a named competitor, what should a company of this size use. Cover the category, the buyer types and the competitive comparisons, then stop adding.
Freeze that set and take a baseline
Run the same prompts before you change anything. Record who was named, in what order, and in what tone. A frozen set is the only way a trend means the market moved rather than your question changing. Vary the set and you lose the comparison you were building.
Read the answers, not just the count
Open the actual text for ten prompts where you were absent and ten where you were named. The reason is usually on the page, in the shape of a competitor's comparison post or a directory listing you are missing from. The count tells you the score. The text tells you the cause.
Fix the description of what you are
Before you chase placements, make one sentence about your category and buyer appear consistently across your site, your profiles and anywhere you are listed. If assistants are placing you in the wrong category, no amount of coverage in the right one will land cleanly.
Go where the answer already goes
Identify the pages that get cited in answers to your prompts. Getting accurately represented there, through listings, submissions, corrections or earned coverage, is the most direct lever available. Correcting an out-of-date entry is often faster than earning a new one.
Rewrite your own pages to answer the question
One page per real question. Open with the answer, then the detail. State the facts plainly rather than implying them. This overlaps heavily with good search practice, which is why it is cheap to do. It is the same discipline with a stricter editor.
Re-run the same set on a weekly rhythm
Weekly is frequent enough to catch a sudden drop and slow enough to avoid reacting to sampling noise. Monthly hides the moment something broke. Daily invites you to over-read variation that would have reverted on its own.
Attribute movement carefully, or not at all
Before you credit a change to your work, check whether the model version behind the answers changed in the same window. If it did, say so in the report and hold the conclusion. Naming the uncertainty costs you nothing and protects the next number you present.

Where effort tends to go

The table below is our reading of the mechanism, not a measured effect. No one outside the providers can isolate a single cause behind a single answer. Treat each row as a hypothesis with a way to test it, and weight your time accordingly.

Moves, mechanisms and how you would know

The move	What it plausibly affects	How you would know
State plainly on your own pages who the product is for and what it does not do	What a retrieval pass can quote, and which category you get sorted into	You start appearing for prompts that name the category and buyer, not only prompts that name you
Get accurately listed on the comparisons, directories and round-ups that already answer your category question	The pool of pages an answer is likely to be written from	Answers cite or paraphrase those pages, and your name appears beside the same competitors
Publish specific facts: prices, limits, integrations, constraints	Whether an assistant has anything concrete to say about you beyond adjectives	Descriptions of you get more accurate and less generic across prompts
Resolve contradictory descriptions of your company across your own properties	Entity clarity and the odds of being confused with a similar name	Fewer answers that describe a different company or blend two together
Publish an llms.txt file	Unproven. We are not aware of a provider publicly committing to read it	Hard to isolate from anything else. A cheap experiment, not a lever
Stuff a page with phrases you hope the assistant matches	Little you want. A paragraph generator is not a keyword matcher	No reliable signal, and a page your buyers will not enjoy reading

The objection you should keep: the answers change

Ask ChatGPT the same question twice and you can get two different lists. That is not a flaw in your measurement. It is a property of the system. Anyone selling you certainty about it is overselling. Keep the objection, then handle it.

Three separate things move a result. Sampling variation, where the same model gives a different answer to the same prompt. Retrieval variation, where the pages reachable at that moment differ. And model updates, where the provider ships a new version and the behaviour shifts underneath everyone at once. Only the third is a step change. It is also the one most likely to be misread as your work succeeding or failing.

The practical consequences are simple. Never read one answer as evidence of anything. Read distributions across a fixed set of prompts, run at a fixed cadence, and compare like with like. Treat a single week's move as a signal to look, not a result to report. When you do report, put the caveat next to the number rather than in a footnote: a week-over-week shift may reflect a model update rather than a real competitive change.

This also sets the ceiling on what any tool in this category, including ours, can promise. A tracker can tell you what a given set of models said to a given set of prompts in a given week. It cannot promise the number is stable. It cannot guarantee a placement. Anyone who says otherwise is describing a system they do not control.

How to measure presence so the number survives scrutiny

A count of mentions is the weakest useful metric. Build up from it.

Mention rate across a fixed set

Of the prompts you froze, how many named you this week. This is the headline, and it only means anything because the set does not change.

Position within the answer

Being named first, as the recommendation, is materially different from being named last, among alternatives. Position within each response is measurable. It often moves before the mention rate does, so it is the earlier warning.

Sentiment, with its limits

Positive, neutral, negative or absent is a coarse but useful cut, especially when a mention exists but carries a caveat. Automated sentiment on short passages is approximate. Use it to find the answers worth reading, not as a score to defend on its own.

Competitors in the same frame

An absolute count has no meaning without the field. If you are named in thirty prompts and the market leader is named in ninety, you know something. If you are named in thirty and nobody else is named at all, you know something very different. Name the competitors up front and measure them with the identical prompt set.

The model version behind every answer

Store it. When the number jumps, the first question is always whether the model changed. Without the version recorded at read time, that question is unanswerable a week later.

What to report, and to whom

This work rarely fails on the merits. It gets dropped because it never reaches a page anyone reads. Decide the reporting shape before you start, because it determines what you need to store.

For the weekly working rhythm, the useful artefact is short. The headline number, the direction, the two or three prompts that changed, and one line on whether a model update could explain it. That is a scan, not a meeting. Anything longer gets skimmed and then skipped.

For the monthly or quarterly deck, the useful artefact is a trend line with the competitive field on it and one paragraph of interpretation. Executives do not need the prompt list. They need to know whether the assistant a growing share of buyers now asks first is naming the company, and whether that is improving. Give them the caveat in the same breath as the number. The first sceptical question in the room will be that the answers change, and you want to have said it before they do.

What to avoid: a single hero number with no field around it, a chart built from a prompt set that quietly changed mid-quarter, and any claim that a specific piece of content caused a specific gain. The evidence does not support that level of attribution. One overclaim will cost you the credibility of everything else on the slide. State what the data shows, name what it cannot show, and leave the editorial call to the people paid to make it.

Questions

How does ChatGPT choose its sources?

The providers do not publish the selection logic, so nobody outside them can tell you precisely. What is observable is that when the assistant searches before answering, it draws on a small set of pages and usually shows them, and those pages tend to be ones that already answer the question directly. Work on being present and accurate in that material rather than on guessing at the ranking function.

Can you optimise for ChatGPT?

You can influence it. You cannot control it. You can change what the public record says about you, how clearly you are described, and whether your own pages give an answer something concrete to quote. You cannot edit the model, buy a position, or guarantee a mention. Anyone promising a placement is promising something they do not own.

Does ChatGPT show which sources it used?

It depends on how the answer was produced. When it searches, it typically shows the links it drew on. When it answers from trained memory, there is nothing to cite and no way to see where the naming came from. This is why a citation audit alone is an incomplete picture of your presence.

How often do ChatGPT answers change?

Enough that a single check tells you very little. Identical prompts can return different lists, retrieval differs moment to moment, and model updates shift behaviour for everyone at once. Measure a fixed prompt set on a fixed cadence and read the trend, not the instance.

Is this the same as SEO?

It overlaps, and it is not identical. Clear, specific, crawlable content helps in both. The difference is the unit of success. Search rewards a ranked link. An assistant rewards being the brand named inside a synthesised paragraph, which puts far more weight on third-party consensus and on how you are described.

Does llms.txt do anything?

We are not aware of a provider publicly committing to read it. It is cheap to publish and its effect is unproven. Treat it as a low-cost experiment rather than a lever, and do not let it displace work on the pages that answers are actually written from.

How do I check whether ChatGPT mentions my brand right now?

Ask it yourself, across several phrasings a buyer would use, and save the answers. That gives you an anecdote, which is a reasonable place to start. It becomes evidence when the same prompts are run repeatedly on a schedule and compared against a named competitive field.

What will Seenby cost?

A weekly digest with drift alerts is part of the plan. The product is not open yet, so the waitlist is the way in.