Which GEO tools do AI assistants recommend? What 1,876 answers showed once I filtered the bad rows

Chris Han, Head of Product at MaxAEO

Ask 8 AI assistants for GEO tools and five names keep coming back: Profound, Semrush, Peec AI, Otterly.AI and Ahrefs. The list was the easy part. Profound's share then went from 77.5% to 83.8% just from me dropping rows, before I'd analyzed anything. If you're building something like this, the rows are what I'd look at first.

I should say I build MaxAEO, a tool that tracks how AI assistants mention brands. It's in the table below, so I'm not neutral. An AI assistant helped me draft this, and a script on our tracker's export did the counting.

The setup: 40 buyer questions, none naming a brand, sent to ChatGPT, Gemini, Perplexity, Claude, Grok, Copilot, Google AI Overviews and Google AI Mode once a day from October 26 to November 1, 2026. That's 2,240 runs.

status: success doesn't mean you got an answer

There were three ways to get no answer:

1. 210 runs failed and reported it.
2. 79 AI Overviews runs were marked success with a blank answer, 78 of them on October 28 to 30.
3. 75 runs were marked success where the "answer" was an error message. 73 of those were Gemini on those same three days.

Every run now goes through this first:

const ERROR_TEXT = /BardErrorInfo|Something went wrong while generating|couldn.t load|Error in message stream/;

function isRealAnswer(run) {
  if (run.status !== "success") return false;
  const text = (run.answer || "").trim();
  if (text.length < 120) return false; // blank, or a one-line stub
  return !ERROR_TEXT.test(text);
}

1,573 answers named Profound, and that count never changed. Only the denominator did: 77.5% over every run marked success, 80.6% with the blanks out, 83.8% with the error text out too. Gemini's own number for Profound went from 60.6% to 85.8%, which is the kind of jump you'd happily write a blog post about if you didn't check.

I also print attempts, errors, blanks and junk per platform per day before reading anything else. A collector outage shows up there as a dead day on two platforms. In a weekly average it can pass for a platform changing its mind.

Don't match brand names by their first word

To work out which tool an answer names first, my script took the first word of each brand and searched for it. SE Ranking became se. The word use contains se. So does search.

Profound's first-place share on ChatGPT came out at 44%. Fixed, it's 49%. Copilot went from 58% to 63%, AI Overviews from 11% to 25%.

The replacement is a fixed list with word boundaries, run on the answer after stripping links:

const BRANDS = [
  ["Profound", /\b(?:Profound|tryprofound)\b/], // case-sensitive: "profound" is a normal word
  ["Semrush", /\bsemrush\b/i],
  ["SE Ranking", /\bse ranking\b|\bse visible\b/i],
  // 44 in all
];

function brandsInOrder(answer) {
  const text = answer
    .replace(/\]\([^)]*\)/g, "]")
    .replace(/https?:\/\/\S+/g, " ");
  return BRANDS
    .map(([name, re]) => [text.search(re), name])
    .filter(([pos]) => pos >= 0)
    .sort((a, b) => a[0] - b[0])
    .map(([, name]) => name);
}

It's still rough. A second pass I wrote separately, matching full names, got 70% on Copilot where this one gets 63%, and 21% on AI Overviews where this one gets 25%. Same direction, a few points apart, so I quote first place as a rough number.

I found a third problem while writing this. Some answers have source labels on their own lines, like HubSpot +1 in Google AI Overviews, and those count as naming HubSpot unless you strip them. It's mostly Google's two surfaces, with some ChatGPT. Re-running October 11 to 27 with those lines removed changed 289 of 5,013 answers. Named rates barely moved (Profound 82.4% to 82.3%). First place on AI Mode moved more, up to 10 points on one tool on a small set of questions. I haven't reworked older numbers for this one yet.

Alias rules decide the ranking

Answers often say "Semrush AI Visibility Toolkit" or "Semrush One" instead of Semrush. Count those as Semrush and it's second at 77.3%. Count only the plain name and it's fourth at 63.9%. Goodie AI goes from 7.0% to 13.8% the same way.

I merged them, because a buyer ends up at the same company. That's a judgment call, so write it down next to the number.

The table

This table comes from the tracker's brand field, which also counts a tool that only appears in a source link. Matching the answer text alone shifts some cells. We drop from 20.0% to 14.6% on AI Mode, and Profound on AI Overviews goes from 72.4% to 75.7%.

Keep platform as its own column. A blended score would hide that Ansvisor is in 44% of Grok's answers and 1% or less everywhere else. YouTube is cited in 73.1% of AI Mode answers and barely anywhere outside Google. 19.4% of ChatGPT's answers cite a Medium post.

Whose first place holds up

Stewart Goodwin (@stridegeo) asked in the comments on my first post which brands keep the top pick best, and what they have in common. I can do the first half. This uses October 11 to 27 and the text matching above: same question, same platform, back-to-back days, 4,366 pairs. The tool named first was still first the next day 42.7% of the time.

Profound held it 56.1% of the time (1,509 pairs), Semrush 44.5%, HubSpot 44.5%, Peec AI 43.1%, Otterly.AI 39.0% and Ahrefs 23.2% (95 pairs). We were at 50.4% on only 125 pairs, so I wouldn't lean on that. Losing first place rarely means vanishing for the top names: Profound was still somewhere in the next day's answer 91.1% of the time.

The second half of his question I can't answer yet. I don't know what the steady ones have in common.

What this doesn't tell you

Which tool is best, for a start. This only counts who gets named. It also ran on our own tracker in a category we compete in. It's one week, and three of those days had collection trouble, so the rows that survived from October 28 to 30 are thinner.

My first post in this series had 41 blank or error rows in its 2,124 answers. I've added a note there.

If you'd rather not build it, this per-platform breakdown is what MaxAEO does.

Written with AI assistance.