Guide · 19 July 2026

Gemini AI and Glotier

Gemini is the assistant most people use without deciding to. It answers inside Google Search, it is the assistant on a billion Android phones, and it sits in the Docs and Gmail tabs that are already open. That makes it the engine most likely to describe your product to someone who has never heard of you, and the one whose answer you are least likely to ever see. This is what it is genuinely good at, what you can do with it, and how to find out what it says about you.

What Gemini actually is now

Gemini is Google’s family of multimodal models, and the word family is doing real work in that sentence. It is not one product. There is the consumer assistant at gemini.google.com, the answers generated inside Google Search, the assistant embedded in Android and Workspace, and the API that developers build on. They share a name and a lineage, and they do not always share a model or a set of capabilities, which is the single most common source of confusion when someone says “Gemini told me”.

As of 19 July 2026, the models Google lists as stable in its own API documentation are these. We are quoting Google’s descriptions rather than paraphrasing them, because this is exactly the sort of detail that the internet gets confidently wrong.

Alongside those sits a preview tier, currently including a 3.1 Pro preview, a 3 Flash preview, live translation and text-to-speech variants, and an omni-flash model. Preview models are where the interesting capabilities appear first and where the interface is least stable.

The part worth internalising is the other end of that list. Google does not quietly leave old models running forever. Gemini 2.0 Flash and 2.0 Flash-Lite are marked shut down, as are a couple of preview models that existed a few months ago. If you are building on the API, pin a specific version, subscribe to the changelog, and assume that a model you hard-coded a year ago will one day return an error instead of a completion. If you are simply using Gemini, the same churn explains why an answer you got last month may be unreproducible today.

What Gemini is genuinely good at

Model benchmarks move every few weeks and are largely useless for deciding what to actually use. Three structural advantages have been stable across Gemini generations, and those are worth planning around.

Long context, and what it is really for

Gemini pushed long context into the mainstream, and the one-million-token window that arrived with the 2.0 generation reset expectations for the whole field. The popular framing is “you can paste a whole book”, which undersells it. The useful version is that you can stop building retrieval pipelines for problems that do not need one.

If your question is about a hundred-page contract, a full codebase, a quarter of support tickets or an entire documentation site, the naive approach of putting all of it in the prompt is now often better than a clever chunking system, because nothing gets lost at the chunk boundary and the model can see relationships that span the whole document. For a small team this is a genuine cost saving: the engineering you do not have to do.

It is not free. Long prompts cost more and respond slower, and models still attend unevenly across a very long context, so burying the decisive sentence in the middle of a million tokens is a poor plan. Treat the window as headroom that removes an architectural problem, not as a licence to stop thinking about what you send.

Native multimodality

Gemini was designed multimodal rather than having vision bolted on afterwards, and in practice that shows up in the unglamorous tasks: reading a screenshot of a dashboard, pulling numbers out of a photographed receipt, working through a diagram, summarising a recorded call. For anyone running a business without a data team, this is the capability that quietly removes the most manual work, because so much of a small company’s information is trapped in images, PDFs and recordings rather than in a database.

Distribution, which matters more than either

This is the one that should change your priorities. ChatGPT is a destination: people choose to go there. Gemini is largely an ambient layer. It answers inside a Google search that someone ran without thinking about AI at all, it is on the phone in their pocket, and it is in the document they already had open.

The consequence for anyone selling something is direct. The person forming a first impression of your product via Gemini is not an AI enthusiast running an experiment. It is an ordinary buyer who typed a question into Google, got a synthesised answer naming three products, and never scrolled. They will not tell you this happened. There is no referral log entry for “decided against you inside an AI Overview”.

What you can actually achieve with Gemini

Setting aside the demos, here is where Gemini earns its place in a small company’s week.

Grounding, and why it is the whole story

Most arguments about AI assistants are about the model. For anyone trying to be recommended, the model is the less important half. What matters is whether the answer was grounded, meaning the assistant went and fetched live pages before writing, or ungrounded, meaning it answered from what it absorbed during training.

The difference is the difference between two completely different games. An ungrounded answer is a memory test, and a small product is not in the memory. No amount of work this quarter changes what a model learned last year, and if you are new or niche, an ungrounded assistant will simply never have heard of you. A grounded answer is a retrieval test, and retrieval is a thing you can influence starting this afternoon, because it depends on pages that exist right now.

This is why the question “does ChatGPT know about my product” is usually the wrong one. The useful question is whether the assistant, when it goes looking, finds you. For a young company the grounded path is the only path, and it is a far more hopeful situation than the alternative: you are not waiting for the next training run, you are competing for retrieval against pages you can go and look at today.

Gemini is interesting here because grounding it against Google Search is not a bolt-on. It is Google’s own index, on Google’s own infrastructure, feeding Google’s own model. Whatever you think of the answers, that is the shortest path between the open web and a synthesised recommendation that exists in the industry, and it is why work that improves your standing in ordinary search shows up here faster than it shows up anywhere else.

Gemini, AI Overviews, and the traffic you never see

The Gemini surface with the widest reach is not the Gemini app. It is the block of generated text at the top of a Google results page. Someone searches, reads a paragraph that names two or three products, and never scrolls to the links. From your side, nothing happened. There is no impression, no click, no referrer, no row in any dashboard. A buyer formed an opinion about your category and you have no record of it.

We watched this from the other side on our own site. Two weeks after our domain went live, Google began generating an overview summarising what we do, which we describe in our 61-day case study. The instructive part was not that it happened. It was that it happened without any signal reaching us. We only found it because we went looking.

That is the strongest practical argument for measuring this rather than assuming. Every other marketing channel you run has a feedback loop: a number goes up or down and you react. Being described by an AI is currently the only significant channel with no feedback loop at all unless you build one. The answers are being given whether you look or not, and the default is not to look.

Getting better answers out of Gemini

Practical, and mostly the opposite of what prompt-engineering threads suggest.

Say what you have, not what you want it to be

Role-play preambles do very little on current models. Context does a great deal. “You are a world-class analyst” changes almost nothing; pasting the actual document, the actual numbers and the actual constraint changes everything. If you have a long-context model available, the highest-leverage prompting technique is simply giving it more of the real material.

Ask for the evidence in the same breath

Requesting the supporting quote alongside each claim converts an answer you have to trust into one you can check in thirty seconds. It also tends to reduce invention, because a claim with no quotable support is more likely to be dropped than dressed up. This single habit is worth more than any prompt template.

Make it disagree with you on purpose

Assistants are agreeable, and an agreeable answer to a question you framed badly is the most expensive output there is. Ask for the strongest case against your plan, or for what would have to be true for it to fail. You are using the model as a critic rather than as a mirror, which is the use it is genuinely good at and the one people reach for least.

Decide once whether reproducibility matters

For an exploratory question, grounding is a feature and variance is fine. For something you will quote in a meeting next month, save the answer and the sources when you get them. A grounded answer is a snapshot of a moving web, and treating it as a stable fact is how a confident number ends up in a deck with nothing behind it.

Gemini, ChatGPT and Perplexity, practically

Not a benchmark table. The functional difference that actually affects you, which is how each one reaches your buyer.

ChatGPT is a destination. People go there deliberately, often for longer, more considered sessions. If someone researches your category there, they are likely to go several questions deep, which means there is room to be discovered on the second or third question even if you missed the first.

Gemini is ambient. It answers where the person already was, most often inside a Google search they did not think of as an AI interaction. The session is shorter and the first answer carries far more weight, because there may not be a second one. Missing the first answer here often means missing the buyer entirely.

Perplexity is search-first. It retrieves for essentially everything and puts its sources in front of the reader rather than behind a citation marker, which makes it the easiest of the three to audit and the most revealing about which pages are winning your category.

The reason we ask all three rather than picking one is that they disagree, and the disagreement is not a rounding error. A product can be named consistently by one assistant and be entirely absent from another, because they read different pages. Checking a single assistant and drawing a conclusion is the most common measurement mistake in this whole subject, and it is an easy one to avoid.

Where Gemini will let you down

A guide that only lists strengths is marketing. Four honest limits.

It is confident when it is wrong, like every model in this class. Grounding reduces this but does not remove it, because a retrieved page can itself be wrong and the model has no way to know. Any workflow where a mistake is expensive needs a human check, and no amount of prompt engineering replaces it.

Answers are not reproducible over time. Grounded answers depend on what was retrievable at that moment, and the model underneath you may be replaced or shut down. If you need a result you can point at next quarter, record the answer when you get it. Do not assume you can ask again.

“Gemini” is not one thing. The answer in Google Search, the answer in the Gemini app and the answer from the API can differ for the same question, because they may run different models with different grounding. When someone reports what Gemini said about you, the first question is which Gemini.

It cannot tell you why it named someone. If you ask a model to explain its own recommendation, you get a plausible story rather than the mechanism. The only way to know what actually fed an answer is to look at what was retrieved.

The part that decides whether it recommends you

Here is the whole thing in one paragraph. When Gemini answers a buying question with grounding on, it retrieves pages and writes from them. Your product is named if it is in those pages and described clearly enough to be repeated. It is absent if it is not. There is no ranking to climb and no submission form.

We measured what those pages look like. We put 40 real software buying questions to ChatGPT, Gemini and Perplexity and recorded every source behind every answer. Reddit appeared in 38 of the 40. YouTube appeared in 35% of them. The classic software review directories, the ones companies pay thousands a year to appear on, accounted for 0.7% of all citations, and four of the five were not cited once.

Two caveats we would rather state than have pointed out. That study merged the three assistants, so it describes what they collectively read rather than Gemini alone, and it is one run on one day. The splitting of it by assistant is the next study we will publish. What the numbers are good for is the shape, and the shape is not close to the margin.

Gemini has one property the other assistants do not, and it changes your strategy. It lives inside Google. The retrieval that feeds a grounded Gemini answer and the index that decides ordinary search results are the same machinery. Which means SEO work is not a separate track from AI visibility here, it is the same track. The page you got ranking last quarter is a page that can now be retrieved and quoted. That overlap is the most under-used advantage available to a small team, because it means the work compounds twice.

Why Glotier and Gemini fit together

Glotier exists because of the gap described above: the answers being given about you are invisible to you. Your analytics show the visits you got, never the recommendation you lost inside somebody else’s chat window.

So we ask. We put your buyers’ real questions to Gemini alongside ChatGPT and Perplexity, as live web queries rather than from training memory, and we record three things: whether you were named, who was named instead, and which pages the answer was built from. That last one is the useful one. Knowing you are absent is a fact. Knowing which four domains decided the answer is an instruction.

Asking all three matters more than it sounds, and Gemini is the reason. Because it answers inside Search and on Android, it reaches people who never opened an AI product on purpose. A founder who checks ChatGPT, sees their name and relaxes has measured one assistant, and quite possibly not the one their customer used. The three disagree often enough that a single check is a coin flip.

The rest of what we found says the same thing in a different way. Across those 40 questions the answers named 266 distinct products, and 223 of them appeared in exactly one question. No product was named in more than three out of forty. Being named once is not a position, it is a draw you happened to win, and the only way to see the difference is to ask many questions and count.

What we will not do is claim to know the mechanism inside Google. Nobody outside Google does, and the companies selling you a “Gemini ranking factor” are guessing. What can be honestly measured is the input and the output: the question, the pages retrieved, and the names that came back. That is what we measure, and it is enough to act on.

A practical checklist

In the order we would do it, with no budget:

Sources and dates

Model names and Google’s descriptions of them were taken from Google’s own Gemini API model documentation on 19 July 2026. We quote it rather than paraphrase because that lineup changes frequently and second-hand summaries of it are unreliable, including several that currently rank well for the obvious searches. If you are reading this months later, check the model page before trusting the table above.

The citation figures come from our own study of 40 software buying questions run on 19 July 2026, described in full, including its limits and the eight-source-per-answer ceiling, in what sources AI actually cites. We have not repeated third-party market share statistics here. The numbers circulating for AI assistant usage vary wildly between sources and we could not verify any of them to a standard we would want quoted back at us.