Skip to content
SilktideHelp

Searches screen

Before an assistant answers a question, it often decides it needs to look something up first. It breaks your prompt into smaller questions and searches for each one. This screen pairs each answer with the searches it ran to get there.

These are the assistant's own words, not yours and not ours. A simple factual prompt may produce a single search. A comparison prompt can produce fifteen.

Why this is not a keyword report

It is tempting to read this as a keyword list and start optimising for the phrases in it. That is the wrong use.

A search here tells you what the model thought it needed to know. It is a window into how your prompt was interpreted, which is a diagnostic, not a target. Chasing these phrases optimises for one assistant's internal reasoning on one day.

Were you a candidate?

The first thing this screen says, in one line at the top, is whether an assistant was ever shopping for you.

An answer's first search is written from memory, before the assistant has read anything. Who it names is therefore the assistant's own shortlist: the brands it had already decided were worth researching. Every search after it is the assistant going to check one of them, usually one vendor at a time.

That split puts each brand you track in one of four positions:

  • Shortlisted and checked - named up front, then looked up directly. The assistant knows you and went to see. What it found decides the rest.
  • Shortlisted, never checked - named up front, never looked up. It knows you and did not think you were worth confirming.
  • Found while researching - not on any first search, but looked up later. Something the assistant read put you in the running; somebody else's page is doing that work.
  • Never considered - on no first search and never looked up.

Never considered is the hardest of the four, and it is not the same as ranking badly. Your pages were never fetched, so nothing on them explains the omission and nothing you change on them is the remedy. The shortlist comes from what the assistant already knows and from the sources it trusts, which is why the finding points at the sites it read instead rather than at your own site.

The line under the verdict counts the products those first searches did name, including ones you do not track, because the number is what separates "the assistant shortlisted nine products and none were you" from "nobody was shortlisted here". It counts names this workspace has resolved a website for, so treat it as a floor.

Each brand's own position is written under its bar in Searched by name, so you can read your rivals' positions as easily as your own.

Searched by name

Above the tables is the finding the tables cannot state on their own: the brands an assistant looked up by name before it answered a question that named nobody.

This matters because an assistant mostly recommends the brands it went looking for. If you ask "what are the best accessibility tools?" and the assistant immediately searches for four vendors by name, those four were its shortlist before it read a single page. A brand missing from every search was passed over earlier than any citation or mention can show.

Only questions that named nobody count here. "Is Silktide worth it?" puts the name into the search itself and proves nothing, so those searches are left out and the panel says how many.

Every brand you track gets a row, in order of how often it was searched for, with its position written underneath, and a brand nobody looked for reads Never - including your own. "The assistant never looked for you" is the finding worth acting on, so it is stated rather than left as a gap. Click a brand to see the searches behind its number.

A brand counts when a search says any of its names, so an alias or a product name finds it too, as does a site: search aimed at its own domain.

Each stretch of words in a search is credited to one name and one name only. Where two names could claim the same words, a brand you track claims them first, and between two names the longer one wins. A search saying "Deque axe Monitor enterprise accessibility" counts once, for Deque, and not a second time for "axe" - the same reading the highlighting and the Who it searched for column show you below, so a number and the searches behind it can never say different things.

If a name you would expect to count is being missed, check Entities: a name that has not been attached to a brand yet is credited to nobody, here and everywhere else in Presence.

Brands, or everything it went looking for

The toggle in the panel's header decides which names are ranked. Brands is the workspace you track, as above. All adds the names Silktide recognises that nobody has attached to a brand, ranked in the same list against the same searches.

The wider list is usually the truer answer to "what was it shopping for?". In one workspace here the assistants searched for two of the brands it tracks, and for twenty-five other products by name - Semrush, Ahrefs, BrightEdge, Conductor - several of them more often than either tracked brand. Restricted to the roster, that panel answers "two" to a question whose answer was twenty-seven. As in the table below, a brand you track carries its icon and an unattached name is outlined, so the two never read as one thing.

An unattached name has to earn its place first, on the same evidence the Who it searched for column asks for and for the same reason: without that bar, words a mention parser mistook for products would outrank every real brand in the list.

One name to a stretch of words applies across both lists, which is what keeps a fragment out of the ranking. "axe" counts where a search says it on its own and not where the search says "Deque axe Monitor", and a name that only ever turns up inside a longer one counts nothing and gets no row at all. A pair like that usually means one product family arrived as two names, which is worth tidying in Entities.

A name nobody searched for has no row at all, where a tracked brand in the same position keeps one reading Never. The two absences say different things. "The assistant never looked for you" is the finding this panel exists to state, while an unattached name that did not come up is one of several hundred names that did not come up, and drawing them all would bury the top of the ranking - the part that says what the assistant wanted. All is a ranking rather than an inventory for the same reason, so it holds the twenty-five busiest names; the table below is where you go to see everything.

The toggle appears only where the workspace has unattached names in scope. A second setting that is always empty is worse than no setting.

Clicking a brand sets the Searched by name filter above the table, which is a normal filter - you can switch brands or clear it without going back. It matches on everything the count matches on, so a search that only said an alias is in the list too, and it leaves out the searches the panel left out. A brand reading Never opens onto an empty list, which is the honest answer rather than a contradiction. Clicking an unattached name does the same thing to the same table, counted the same way, so a row's number and the list it opens can only ever agree.

The shortlist over time

Beside the bars is the same finding as a trend: how often each brand was on the assistant's shortlist, period by period. Whether you are a candidate at all is the one thing on this screen worth watching rather than reading once, and it is a line no citation report can draw.

Two views. Share is the percentage of answers that revealed their searches where the brand was on the first search, and it is the honest comparison over time, because how many answers reveal their searches varies day to day. Answers is the count underneath it, for when a percentage resting on eighteen answers should not read the same as one resting on two hundred. The chart's data table always shows the counts, whichever way the lines are drawn.

The chart will not draw a period as a zero unless it measured one. A period nothing revealed a search in is a gap in the line, not a floor: assistants do not report their searches every time, and an unreported day drawn at the axis invents a collapse that never happened. Today is left out for the same reason until it finishes, since a running total falls off a cliff every morning. But a period that did reveal searches and never named you is drawn flat on the axis, because that is a real zero and the finding this screen exists to state.

Searches cannot be backfilled, so the trend starts from the first batch that captured them. Until two periods have been measured the chart says so rather than drawing a line through one point.

The table: one answer, and what it went looking for

Each row is one answer: the question it was given, the searches it ran before writing a word, and the date it ran them. Click the row to open the answer itself.

The pairing of question and searches is the point. A search on its own is close to gibberish; next to the question that caused it, it shows you what the assistant thinks your customer's question is really about.

The searches read in the order the assistant ran them, and each is labelled with what it was for:

  • Shortlist - the answer's first search, written from memory. "best enterprise accessibility platforms level access deque siteimprove 2026". This is where the assistant decides who the candidates are.
  • Checks name - a later search reaching for one vendor, often ending in the word "official" as it goes looking for that vendor's own page.
  • Reads site - a site: search, aimed straight at one site's pages.
  • Follow-up - a later search naming nobody in particular, or several names at once.

The labels come from the data, not from a guess. Across two workspaces, "best" and "2026" appear in twenty-four first searches and no later ones, while "official" appears in forty-one later searches and three first ones; first searches average eleven words and later ones six and a half. The assistant recalls a shortlist, then works down it.

Where a search says a name we recognise, that name is highlighted in place. The words stay as the assistant wrote them, because the word it chose is itself a finding: a search saying "Monsido" is looking for Acquia, and an assistant a rename behind is worth knowing about. Who it searched for then names who those words resolve to, once each, however many searches named them.

That column is wider than the brands you track, deliberately. It also lists names Silktide has seen in answers but nobody has attached to a brand - Semrush, Webflow, Hotjar, Drupal. Restricted to tracked brands the column is empty on most rows and reads as broken, when the truth is that the assistant was shopping for real products you have not told us to follow. Brands you track carry their icon; the rest are outlined, so the two never read as one thing.

A name has to earn its place: at least three answers must have put it forward as an answer in their own right. Without that bar, words a mention parser mistook for products - "Analytics", "Enterprise", "Easy Checks" - would outrank every real brand here.

Use the arrow beside a search to re-run it in Bing, which is the index ChatGPT's browsing reads. It is close to what the assistant was ranking, not identical: its results were personalised and timestamped, and OpenAI has been moving toward an index of its own.

A question marked Branded named a brand itself. Those rows prove less: an assistant looking up your rival after you asked about your rival is arithmetic, not a shortlist.

Asking the same question twice makes two rows, which is why the date is there and why the table is sorted by it. Re-runs are worth comparing: the same question usually produces different words for the same intent, and a brand that appears in one run and not the other is a brand the assistant is unsure about.

The Sites tab is a second, sharper table. It lists the domains an assistant aimed a site: operator at, meaning it decided - before it had an answer - that this particular site was worth going to directly.

That is not the same as a citation. A citation says a page ended up in the answer. A named-site search says the model went looking there on purpose, whether or not it used what it found.

What is not here

Only some assistants report their searches. ChatGPT does; Gemini does not report them at all. An assistant that reports nothing contributes no rows rather than empty ones.

Searches are recorded as answers arrive, and cannot be backfilled. Existing workspaces show nothing here until their next batch runs.

Last updated

Was this page helpful?

Searches screen | Silktide Help