The Impact of Meta Titles on AI Search Visibility

Experiment

We rewrote 7 meta titles to match the target prompt, and tracked them across ChatGPT, AI Overviews, and Perplexity. Retrievals shifted. Brand mentions didn’t.

Jordan Parkes
Experiment Gauge Comparison

Key Findings:

  • Changing the meta title did not improve AI visibility, but it didn’t hurt it either.

  • Retrievals shifted toward the new title on ChatGPT and AI Overviews: relative to control trends, target-prompt retrievals rose ×2.75 and ×2.64, while Perplexity did not react.

  • The title moved retrievals rather than adding them: page-level ChatGPT retrievals stayed flat at 121, 119, and 122.

  • More retrievals did not mean more mentions: of four brands whose pages were retrieved more often for the target prompt, only one gained visibility.

  • A meta title change only helped pages that were already retrieved: For three of seven brands, ChatGPT and AI Overviews seldom retrieved the test page.

  • AI visibility fluctuates heavily on its own: control prompts fell 19.3 points on average with no title change, so short-term swings of 10 to 20 points are normal.

Executive Summary: ZeroClick Labs found that rewriting a page’s meta title to match a target prompt did not increase how often AI search tools mentioned the brand, but it did change how often the page was retrieved for that prompt. The title appears to influence which sources an AI tool selects, but not which brands it recommends.

 

Today, AI-generated answers often stand between searchers and the websites they used to visit. They also heavily influence buyers’ perception of a brand, as well as their choices: 85% of software buyers think more highly of a vendor included in an AI answer, and 69% have chosen a different vendor than planned because of an AI recommendation.

That has made being named inside an AI answer a goal in its own right for many brands, and there’s a host of AI optimization strategies to make that happen. One of those is rewriting a page’s meta title so it matches the prompt the brand wants to win.

In traditional search, the title element has a documented role: Google uses it, among other sources, to build the title link shown in results, which Google calls “often the primary piece of information people use to decide which result to click.”

In AI search, evidence that a title change alone moves the needle has so far been mostly anecdotal. This study tests whether changing only a page’s meta title affects how often ChatGPT, AI Overviews, and Perplexity mention the brand or retrieve the page.

The claim we tested

The argument: AI crawlers discard much of a page’s metadata, but the title element usually survives.

  • If the title is one of the few page-level signals a model reliably sees, matching it to a prompt should improve the brand’s chances of being mentioned.

We designed this study to test the above claim under controlled conditions. We measured visibility, meaning how often AI answers mention the brand, and retrieval, meaning how often the AI tool fetches the page. We did not examine crawler behavior directly.

Hypotheses

The results are graded against four hypotheses:

Table 1: Hypotheses and what each would look like
HypothesisWhat it would look like
+2Target prompts improve for most brands, and the previous-title prompt declines
+1Target prompts improve for most brands, and the previous-title prompt is unaffected
0No measurable change in target prompts relative to the control prompt
-1Target prompts decline measurably relative to the control prompt

Methodology (overview)

  • Sample: 7 pages from 7 brands (one national software company, one regional and five single-metro service businesses), each retitled on 21 September 2026 to match a target prompt.
  • Prompts: Each page was tracked against 3 prompts: the target prompt, a related control prompt that never appeared in any title, and the prompt its old title was built around. Each brand also had one unchanged control page.
  • Platforms: ChatGPT and Google AI Overviews, with Perplexity tracked alongside, via Peec AI.
  • Period: A two-week baseline (7 to 20 September) and four weeks after the change (21 September to 18 October), with about 14 runs per prompt per platform in each two-week window.
  • Metrics: visibility (the share of answers that mention the brand) and retrieval (how often the AI tool fetched the page).
  • Statistics: retrieval effects are reported as multipliers relative to the control trend (×1 = no change; ×2.75 = 2.75 times the expected retrievals). The p-value is the probability of seeing an effect at least that large by chance if titles had no effect; values below 0.05 are treated as statistically significant.
  • Full design and limitations, metrics, statistical approach, and data quality notes are in  Appendices A to D.

Do meta titles influence AI visibility?

  • Finding 1.1:  Changing the meta title neither helps nor hurts AI visibility.

For each brand, we tracked how often AI answers named the brand for three types of prompts: target (the prompt the new title was written to match), control (a related prompt that was never in the title), and previous-title (the prompt the previous title used to match). Over the four weeks after the title change, visibility fell for all three prompt types.

Target and control prompts fell by about the same amount

Average visibility across seven brands, all three platforms combined. Target -20.3 points, control -19.3 points, previous-title -12.1 points, baseline to weeks 3 and 4.

  • Target prompt
  • Control prompt
  • Previous-title prompt
BaselineWeeks 1 and 2Weeks 3 and 4Target prompt 28.6%48.9%Control prompt 37.9%57.1%Previous-title prompt 32.7%44.9%

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep – 18 Oct 2026 · changes calculated from unrounded values

As the chart shows, the target prompt went from being named in about half of answers (48.9%) to under a third (28.6%). The control prompt declined by about the same number of points as the target, even though it was never in any title, old or new. If the new title had helped, the target prompt should've done better than the control - which it didn't. 

This suggests that the decline is not specific to the title change, but rather the result of outside factors affecting every prompt at once (e.g., AI tools changing their answers, or normal fluctuation in which brands they name).

In relative terms (as a share of where each prompt started), the same figures are:

  • Target prompts: 42% decline
  • Control prompts: 34% decline
  • Previous-title prompts: 27% decline

The target prompt's decline looks bigger here mostly because it started lower, so a similar drop in points makes up a larger share of it.

Brand by brand, the title made no consistent difference

To account for each brand's own conditions, we compared each target prompt's change with the change in that brand's control prompt. A positive value means the target prompt did better than the control prompt.

No consistent gain over the control prompt

3 of 7 brands ahead of their control prompt after four weeks. Target-prompt change minus control-prompt change, vs baseline, all three platforms combined.

  • Ahead of control
  • Behind control
  • Weeks 1 and 2 (lighter)
  • Weeks 3 and 4
-30-20-10+10+20+300Percentage points vs control prompt (0 = same change as the control)Brand A+5+2Brand B-25-33Brand C+3+15Brand D+15+23Brand E-8-3Brand F-19-9Brand G-1-2

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep - 18 Oct 2026

The same three brands (A, C, and D) finished ahead of their control prompt in both periods, and the other four finished behind. Averaged across all seven, target prompts ended 1.0 point behind the control prompt after four weeks.

  • Brand D finished furthest ahead (+23 points), but its target prompt did not improve. It fell from 19% to 3%, while its control prompt fell further, from 82% to 43%.
  • Brand C (+15 points) is the only brand whose target prompt rose over four weeks, from 38% to 41%, while its control prompt fell 12 points.
  • Brand A (+2 points) stayed roughly level with its control prompt. Both declined, with the target falling from 24% to 19%.
  • Brand B trailed by the widest margin (-33 points). Its target prompt fell from 84% to 50%, while its control prompt ended where it started (71% at baseline, 81% in weeks one and two, 70% in weeks three and four).

Other comparisons point the same way

Against the average of the control prompt and the previous-title prompt, only Brand A finished ahead after four weeks (+6.5 points), and only Brand D beat that average by 10 points or more in either window (+16 points in weeks one and two).

In addition, we graded each brand against the hypotheses on each platform. To count as a gain, the target prompt had to beat both the control and the previous-title prompt by at least 10 points (see Appendix B):

  • ChatGPT: six brands at 0 and one at -1 (Brand F). Brand A's 0 means ChatGPT never mentioned the brand for any prompt in any window, so there was nothing for the title to change.
  • AI Overviews: four brands at 0, two at -1 (Brands B and E), and one at +1 (Brand C). Brand C's AI Overviews visibility rose from 0% to 33%, but on only 6 to 12 Overview appearances per window. Brand F's grade relies on a previous-title prompt with no Overviews at baseline. Against its control prompt alone, it would be +1, although all three of its prompts collapsed in weeks three and four.
  • Perplexity: five brands at 0, one at +2 (Brand A), and one at -1 (Brand G).

Most brand-platform results show no measurable change

15 of 21 graded 0, 4 graded -1, 2 graded above 0. To count as a gain, the target prompt had to beat both the control and the previous-title prompt by at least 10 points.

  • Gain (+1 or +2)
  • No measurable change (0)
  • Decline (-1)
ChatGPTAI OverviewsPerplexityBrand A0*0+2Brand B0-10Brand C0+1**0Brand D000Brand E0-10Brand F-10***0Brand G00-1
  • * Brand A, ChatGPT: never mentioned for any prompt, so 0 means there was nothing to change.
  • ** Brand C, AI Overviews: based on only 6 to 12 Overview appearances per window.
  • *** Brand F, AI Overviews: would be +1 against the control prompt alone.

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep - 18 Oct 2026

Whichever comparison we use, the conclusion remains the same: changing the meta title produces no consistent gain in AI search visibility.

The previous-title prompt did not lose ground

  • Finding 1.2: Removing a prompt from the title did not leave it measurably worse off after four weeks.

Hypothesis +2 predicted that the prompt removed from the title would decline relative to the control prompt. In weeks one and two, there was a small move in that direction: the previous-title prompt fell 8.0 points against 4.0 for the control prompt, and five of seven brands trailed their control.

However, by weeks three and four, the pattern had reversed. The previous-title prompt was down 12.1 points against 19.3 for the control prompt, and only three brands (A, B, and E) trailed their control. For Brands C and D, the previous-title prompt gained visibility, rising from 5% to 29% and from 33% to 45%. 

One brand shows a likely tracking anomaly

Brand F's visibility fell to near zero on all three prompts in weeks three and four: 6% for the target, and 0% for both the control prompt and the previous-title prompt. Over the same weeks, AI tools kept fetching its page: 18 target-prompt retrievals across all three platforms, and 14 Perplexity retrievals for the previous-title prompt. 

A brand that genuinely drops out of AI answers usually loses retrievals too. Mentions vanishing while retrievals continue points to a measurement issue, such as the brand's name no longer being recognized, rather than a real change or an effect of the title. Even so, excluding Brand F does not break the pattern: target prompts fell 15.0 points, the control prompt 15.3 points, and the previous-title prompt 7.7 points.

Do meta title changes influence AI search retrieval?

  • Finding 2.1: Pages with updated meta titles continued to be retrieved at the same rate for the prompts aligned with their revised titles. 

Meta title changes did not increase brand mentions, but they influenced how AI search tools retrieved and used the updated pages. Four weeks after the change, the edited pages were retrieved as often for target prompts as they were before. Every other prompt type, on both the edited and the unchanged pages, lost a large share of its retrievals.

Table 2: Retrievals across ChatGPT, AI Overviews, and Perplexity, summed over all seven brands
Prompt typeTest pagesChangeControl pagesChange
Target prompt64, 76, 640%43, 36, 19-56%
Control prompt62, 51, 44-29%80, 68, 47-41%
Previous-title prompt108, 93, 70-35%31, 26, 16-48%
All study prompts234, 220, 178-24%154, 130, 82-47%
*Each cell lists the baseline, weeks one and two, and weeks three and four. Change compares weeks three and four with the baseline.

If the title change had no effect on retrieval, the test pages should have followed the same path on the target prompt as they did on the control prompt, or as the control pages did. Instead, target-prompt retrievals for the edited pages held level while the other prompt types fell by 29% to 56%, suggesting the new title kept AI tools fetching the page for the prompt it now matched, even as retrievals fell everywhere else. 

Note: The effect comes from ChatGPT and AI Overviews; Perplexity showed no shift.

Target-prompt retrievals doubled on ChatGPT and AI Overviews

The pattern is clearer on the study's two focus platforms. Target-prompt retrievals for the edited pages roughly doubled after the change and stayed up through weeks three and four. On AI Overviews, raw retrievals did slip in weeks three and four (10 to 6), but all AI Overviews retrievals fell in those weeks, so the target prompt still ran well ahead of the trend (×3.36, p = 0.008).

Table 3: ChatGPT and AI Overviews retrievals by prompt type
Prompt typeTest pagesControl pages
Target prompt11, 26, 2313, 17, 13
Control prompt22, 19, 1437, 32, 25
Previous-title prompt38, 28, 1416, 12, 11

Target-prompt retrievals doubled; brand visibility fell with the controls

Indexed to baseline (7-20 Sep) = 100. Test pages, pooled across seven brands. End labels show raw values.

  • Target prompt
  • Control prompt
  • Previous-title prompt
  • Baseline level (100)
Brand visibility, all platforms0100200BaselineWeeks 1-2Weeks 3-4Previous-title 32.7%Control 37.9%Target 28.6%Test-page retrievals, ChatGPT + AI Overviews0100200BaselineWeeks 1-2Weeks 3-4Target 23Control 14Previous-title 14

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep - 18 Oct 2026

The control pages saw a small, temporary rise on target prompts in weeks one and two, then returned to baseline. Most of that rise came from Brand G, whose homepage was retrieved more often alongside its test page. Excluding Brand G, control page retrievals for target prompts went 11, 12, 12, while test page retrievals went 9, 17, 19.

The previous-title prompt moved the other way

On ChatGPT and AI Overviews, the prompt that was removed from the title lost 63% of its test-page retrievals, falling from 38 to 14. That is the largest drop of any group. Over the same period, the control prompt fell 36% on the test pages, and the previous-title prompt fell 31% on the control pages.

On ChatGPT, retrieval follows the title in both directions: toward the new target prompt and, by weeks three and four, away from the old one (×0.34, p = 0.04). On AI Overviews, it moved toward the target prompt, but the previous-title prompt fell no more than the control prompt.

Across all three platforms, the picture is less clear. The previous-title prompt fell 35% on the test pages and 48% on the control pages. On the test pages, Perplexity's previous-title retrievals fell only 20% (70 to 56), in line with its control prompt (-25%). The two-way shift is therefore a ChatGPT result, not a general one.

The shift is significant on ChatGPT and AI Overviews, but not on Perplexity

When we measure each platform on its own, the shift is clear and statistically significant on ChatGPT and AI Overviews, while being absent on Perplexity.

Relative to control trends, the edited pages were retrieved for the target prompt 2.75 times as often on ChatGPT (95% CI 1.35–5.60, p = 0.005) and 2.64 times as often on AI Overviews (1.45–4.81, p = 0.002). On Perplexity, the multiplier was 1.11 (0.63–1.94, p = 0.72), practically indistinguishable from no change.

ChatGPT and AI Overviews retrieved retitled pages ~2.7× more; Perplexity didn't react

Test-page retrievals after the change vs each prompt's baseline, net of the shared trend. Dot = estimate, line = 95% confidence interval.

  • Target prompt (new title)
  • Previous-title prompt
  • No change (×1)
×0.25×0.5×1×2×4Multiplier vs control trend (log scale) · ×1 = no changeChatGPTAI OverviewsPerplexity×2.75×0.58×2.64×0.82×1.11×1.12

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep - 18 Oct 2026 · Poisson regression with prompt fixed effects, post-change windows pooled, robust standard errors

Averaged over both post-change windows, the previous-title prompt fell to 0.58 of its expected ChatGPT retrievals (0.30–1.11, p = 0.10). It showed no detectable change on AI Overviews (0.82, p = 0.77) or Perplexity (1.12, p = 0.69). On retrieval, ChatGPT leans toward the +2 pattern and AI Overviews the +1 pattern.

Perplexity retrieves these pages heavily, about 72% of all page-level retrievals, and kept doing so regardless of the title, suggesting that Perplexity's candidate selection leans more on page content and its own index than on the title string.

Overall, the title change worked as a retrieval signal on ChatGPT and AI Overviews, but not on Perplexity. On ChatGPT, it shifted retrievals toward the new wording and, by weeks three and four, away from the old. On AI Overviews, it added retrievals for the new wording without measurably costing the old. 

On this evidence, a title rewrite looks like a ChatGPT and AI Overviews optimization lever, and its trade-off, losing retrievals for the old wording, shows up only on ChatGPT.

A strong signal, not a precise estimate

This is the most consistent result in the study, as it held in both post-change windows and against both kinds of control. Still, it should be treated as a strong directional signal, not a precise estimate because, while titles very likely shift retrieval on ChatGPT and AI Overviews, pinning down the size of that shift needs a much larger sample.

The combined ChatGPT and AI Overviews effect (×2.97, p < 0.001) stays significant whichever brand is left out. The caveats are scale and concentration. Counts are small - they come mostly from four brands (B, E, F, and G), and control-page volumes vary widely. On its own, the ChatGPT result depends on Brand B, and the AI Overviews result weakens without Brand E or F.

Subscribe for practical updates on AI discovery, answer engine visibility, and the shifts that matter most for brands trying to stay visible.

Do more AI search retrievals lead to more brand mentions?

  • Finding 3.1: Title-driven retrieval gains did not lead to more mentions.

A page being fetched by an AI tool is not the same as the brand being named in the answer. Four brands saw their edited page retrieved more often for the target prompt on ChatGPT or AI Overviews, but only one of them (Brand G) gained visibility on either platform. Measured against its own controls, only Brand B clearly lost ground, and only on AI Overviews.

Table 4: Brands whose target-prompt retrievals rose on ChatGPT or AI Overviews
BrandTarget retrievals (ChatGPT and AI Overviews)ChatGPT visibilityAI Overviews visibility
Brand B5, 6, 9100%, 93%, 100%100%, 71%, 30%
Brand E1, 2, 293%, 79%, 36%100%, 33%, 0%
Brand F3, 9, 871%, 36%, 7%60%, 40%, 14%
Brand G2, 9, 436%, 57%, 43%13%, 10%, 0%
*Each cell lists the baseline, weeks one and two, and weeks three and four. AI Overviews visibility is based only on days an Overview appeared. For Brand E, that was about 2, 3, and 3 days, so its AI Overviews figures are indicative only. Brand F's figures for weeks three and four are affected by the likely tracking anomaly described in Finding 1.

Brand B held its ChatGPT visibility at 100% but lost most of its AI Overviews visibility despite more retrievals. Brand G is the only case where ChatGPT visibility rose with retrievals, but its unchanged homepage also gained retrievals, and its previous-title prompt's ChatGPT visibility rose too (21% to 43%) - implying that something other than the title lifted Brand G on ChatGPT.

Why retrieval and mention can diverge

One explanation is that the title helps a page get selected as a candidate source, while the decision to name a brand depends on other inputs, which may include the page's content and how easily it can be extracted, the brand's third-party coverage, and the competitors that appear in the same results.

The data shows how loose the link is on each platform:

  • Across all prompts and windows, ChatGPT visibility tracks the level of retrievals (Spearman ρ = 0.59). But a change in a prompt's ChatGPT retrievals says almost nothing about the change in its ChatGPT visibility (ρ = 0.10, p = 0.52). ChatGPT names brands it already associates with the query, and extra fetches of the brand's own page did not shift that within four weeks.
  • On AI Overviews (ρ = 0.42) and Perplexity (ρ = 0.46), changes in retrieval and visibility appear to move together, but both figures are inflated by data issues. On AI Overviews, a prompt with no Overview scores zero on both measures, which builds in a correlation; without those cases it drops to ρ = 0.23 (p = 0.16).
  • On Perplexity, the link shows up only in weeks three and four, when visibility dropped to exactly 0% on 9 of 21 prompts (see Appendix D).

The gap also appears in the tracking platform's own research. Peec AI's February 2026 data shows that 64% of URLs Perplexity retrieved were never cited, and their July 2026 analysis found one brand that ranked in the top five by retrieval share but only sixth in brand mentions. 

This fits other published work, such as the research by SparkToro and Gumshoe.ai, which found that individual AI answers vary from run to run, but the brands they draw from tend to come from a fairly stable consideration set. In this process, a title change can alter which page represents a brand, but not whether said brand necessarily makes the set.

The practical point is that retrievals are not a reliable stand-in for visibility, and a report that treats them as the same thing can show progress where brand mentions are flat or falling.

How did the edited pages perform?

  • Finding 4.1: Page-level retrievals fell slightly, mostly on Perplexity.
  • Finding 4.2: Platforms moved differently, often in opposite directions.

The edited pages performed slightly worse overall, and differently on each platform. Across everything the tracking platform monitors, the edited pages were retrieved a little less after the change (mostly because of Perplexity), and the title change doesn't explain the biggest swings. 

What the title did change was where the pages were used: on ChatGPT, they were fetched just as often overall but more often for the new wording. Visibility fell on all three platforms, and individual prompts often moved in opposite directions from one platform to the next, so a single blended score can hide what's really happening.

Overall performance

Total retrievals of the seven test pages went from 692 at baseline to 688 and then 627, a 9% decline. The average citation rate fell from 0.66 to 0.63 and then 0.44, meaning that when AI tools did fetch the pages, they cited them less often.

Table 5: Page-level retrievals and citation rate by brand
BrandRetrievalsCitation rate
Brand A104, 96, 400.2, 0.2, 0.0
Brand B121, 90, 1240.5, 0.4, 0.2
Brand C66, 101, 1300.3, 0.5, 0.2
Brand D102, 76, 590.4, 0.4, 0.2
Brand E211, 236, 1991.0, 0.9, 0.8
Brand F53, 59, 421.1, 0.8, 0.6
Brand G35, 30, 331.1, 1.2, 1.1
Each cell lists the baseline, weeks one and two, and weeks three and four.

Perplexity accounts for about three-quarters (72%) of page-level retrievals, so it drives most of these totals. The two largest swings both came from Perplexity: Brand A fell from 104 to 40, and Brand C rose from 66 to 130. Neither swing is explained by the title change: Brand A's fall hit all three of its prompts, and Brand C's target-prompt Perplexity visibility ended where it started (86%).

Two brands show the range of outcomes at page level:

  • Brand E is a possible downside case. Its page-level ChatGPT retrievals fell from 16 to 8 to 3, and its ChatGPT citation rate from 1.4 to 0.9 to 0.0. Changing the title from "installers" to "companies" may have cost the page ChatGPT retrievals for other queries.
  • Brand G gained a new source of retrievals. Its test page had no AI Overviews retrievals at baseline, then 5 and 9 after the change. However, with only one brand, it's impossible to tell whether that was due to the title or ordinary changes in Google's index.

Per-platform performance

Page-level retrievals moved differently on each platform, but target-prompt visibility fell on all three, with AI Overviews falling furthest:

  • ChatGPT: page-level retrievals held flat (121, 119, and 122) while target-prompt retrievals nearly tripled. The title changed which prompts the pages were fetched for, not how often.
  • AI Overviews: page-level retrievals fell by a third (68, 79, and 46), mostly because of Brand E, so the target prompt held up while the pages lost ground overall.
  • Perplexity: page-level retrievals fell 9% (503 to 459).

Against the control prompt, target prompts fell further on ChatGPT (by 1.9 points) and Perplexity (by 6.7 points), and less far on AI Overviews (by 5.5 points). No platform shows a consistent advantage for the target prompt.

Target-prompt visibility fell on every platform, furthest on AI Overviews

Target change, baseline to weeks 3 and 4: ChatGPT -17.3, AI Overviews -30.1, Perplexity -22.4 points. Average visibility across seven brands.

  • Target prompt
  • Control prompt
  • Previous-title prompt
0%20%40%60%80%**ChatGPTBaselineWeeks 1-2Weeks 3-443.9%26.6%AI OverviewsBaselineWeeks 1-2Weeks 3-450.7%20.6%PerplexityBaselineWeeks 1-2Weeks 3-451.0%28.6%*
  • * The AI Overviews previous-title rise is an artefact: Brands E and F had no Overview at baseline and count as 0%. Across the other five brands it fell steadily: 25.8%, 23.8%, and 19.2%.
  • ** Perplexity figures for weeks 3 and 4 are affected by a tracking anomaly.

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep - 18 Oct 2026

Individual prompts often moved in opposite directions on different platforms over the same weeks. Brand E's target prompt fell from 93% to 36% on ChatGPT while rising from 71% to 93% on Perplexity. Brand C's target prompt rose from 0% to 33% on AI Overviews while barely registering on ChatGPT.

This echoes other research. Even within Google, an Ahrefs study of 540,000 query pairs found that AI Mode and AI Overviews cited the same URLs only 13.7% of the time for the same query.

A single visibility score that blends platforms can hide movements that cancel each other out.

How much does AI visibility fluctuate on its own?

  • Finding 5.1: AI visibility moves a lot on its own.

Even with nothing done to them, control prompts lost an average of 19.3 points of visibility over four weeks, or 15.3 points excluding Brand F's likely tracking anomaly. The declines by brand ranged from 1 point (Brand B) to 39 points (Brand D, from 82% to 43%), again leaving aside Brand F. Control prompts never appeared in any title, so the study was not designed to affect them, although a title change could spill over to related prompts.

Control prompts lost visibility with nothing changed

Average change -19.3 points across 7 brands (-15.3 excluding Brand F), baseline to weeks 3 and 4. Control-prompt visibility, all three platforms combined.

  • Control-prompt change
  • Likely tracking anomaly
  • Average across all seven brands
Typical swing: 10 to 20 pointsBrand D-39Brand E-22Brand C-12Brand G-11Brand A-7Brand B-1Brand F-43likely tracking anomalyAverage -19.30

Source: ZeroClick Labs Meta Titles Study, tracked in Peec AI · 7 brands · 7 Sep - 18 Oct 2026

For anyone reporting AI visibility to clients or leadership, this is a useful reference point: a change of 10 to 20 points over a few weeks can happen with no intervention at all, meaning that a move of that size is normal noise - it proves nothing without a control to compare it against.

Why the numbers are noisy

AI answers are not fixed - the same prompt can return a different set of brands each time it runs. In the SparkToro and Gumshoe.ai study, researchers ran 2,961 prompts across ChatGPT, Claude, and Google's AI Overviews and found less than a 1-in-100 chance that ChatGPT or Google's AI would return the same list of brands twice for the same prompt.

That research also found that the share of responses that mention a brand (i.e., visibility) is more stable than ranking position, but it needs many runs to settle. For its brand-recommendation tests, the study ran each of 12 prompts 60 to 100 times per platform. This study ran each prompt about 14 times per platform in each window, so a single response changing from mentioned to not mentioned moves a score by about 7 points.

For a single prompt, 14 runs per window means a change of roughly 43 points is needed before it is significant at p < 0.05 (see Appendix C). Most of the prompt-level movements in this study, including many quoted above, are smaller than that. More broadly, gaps of 5 to 7 points between AI visibility scores are often statistical ties.

Verdict against the hypotheses

The visibility results support hypothesis 0. Rewriting meta titles to match target prompts did not improve brand visibility in AI search over four weeks. Target prompts declined by about the same number of points as a control prompt that was never in any title, and three of seven brands finished ahead of their control. Prompts removed from the title lost no more visibility than the control prompt, although on ChatGPT the edited pages were retrieved for them far less often.

Table 6: Visibility results against each hypothesis
HypothesisResultEvidence
+2Not supportedTarget prompts did not improve relative to the control prompt, and the previous-title prompt declined less than the control prompt
+1Not supportedOnly three of seven target prompts beat their control prompt, and none beat both controls by 10 points or more after four weeks.
0SupportedTarget prompts fell 20.3 points and the control prompt 19.3 points, with three brands ahead of their control and four behind
-1Not supported as a title effectTarget prompts declined, but the control prompt declined by a similar amount

In absolute terms, target visibility did fall, which on its own could look like hypothesis -1. However, the control prompt (which never appeared in any title) dropped by a similar amount, meaning that the decline cannot be attributed to the title change.

The title change did have a measurable effect on retrieval on ChatGPT and AI Overviews, where target-prompt retrievals roughly doubled while retrievals for other prompts fell. The gain held on ChatGPT through week four; on AI Overviews it fell back in raw terms but stayed well ahead of the trend. Perplexity showed no detectable change.

Table 7: Retrieval results by platform
PlatformRetrieval patternTarget promptPrevious-title prompt
ChatGPTHypothesis +2×2.75 (p = 0.005)×0.58 (p = 0.10)
AI OverviewsHypothesis +1×2.64 (p = 0.002)×0.82 (p = 0.77)
PerplexityHypothesis 0×1.11 (p = 0.72)×1.12 (p = 0.69)
*Retrieval patterns are based on pooled models. Only four of seven brands had any ChatGPT or AI Overviews retrievals, so they cannot be read as "most brands".

The gap between those two results is the central finding. A title can influence whether a page is selected as a source for a prompt, but in this study that did not measurably change whether the brand was named in the answer. Future testing could look at whether other changes, on the page or elsewhere, convert that retrieval into visibility.

A title rewrite on its own is unlikely to produce a measurable gain in AI visibility, but it may still influence which of your pages an AI tool selects for a prompt. Therefore, write titles that describe the page accurately, use them to steer which page gets fetched for the prompts you care about, and measure retrieval and visibility separately against a control before you take credit - or blame - for any result.

1. Check that the page is already retrieved

A title can only steer retrieval for a page the AI tool already fetches, so confirm the page appears in retrievals first. If it doesn't, indexing, internal links, and content coverage come before the title.

2. Write the title for accuracy, not for mentions

Clear titles that describe the page well still matter for traditional search. Although this study suggests they may also help the right page get retrieved for the right prompts, you should treat that as a supporting benefit rather than a route to more brand mentions.

3. Budget for the prompt you're giving up

On ChatGPT, the prompt removed from the title lost 62% of its retrievals within four weeks. Retitle away from wording only if you can afford to lose it, or make sure another page covers it.

4. Set up controls before you change anything

Visibility for prompts no one touched fell by 19 points on average in four weeks. Without control prompts, that decline could easily be mistaken for a penalty, and a smaller decline could be mistaken for success. Track at least one related prompt you have not optimized, and an unchanged page on the same site, and compare against both.

5. Track retrieval and visibility separately

In this study, retrievals for target prompts held steady while visibility fell by 20 points. Any report that presents rising retrievals or citations as visibility growth can give the wrong picture. Track brand mentions directly and show them alongside retrieval data, not in place of it. 

Do the same for platforms: ChatGPT, AI Overviews, and Perplexity often moved in different directions for the same prompt in the same weeks, so report by platform first and combine only when the platforms agree.

6. Check the sample size behind each number

With around 14 runs per prompt, one response moves a score by about 7 points. Before reporting a weekly or fortnightly change, check how many runs it is based on, as larger samples per prompt make smaller changes interpretable.

7. Look beyond the page for visibility gains

If a page is being retrieved but the brand is not being named, the deciding factors are probably off the page. Third-party coverage, reviews, directory listings, and how the brand is described across the web are more likely to affect whether a model recommends it than a single on-page edit.

1. Study design

The test: For each of seven brands, we rewrote one page’s meta title to closely match a chosen target prompt and changed nothing else on the page. Each test page was tracked against three prompts, and each brand had a second, unchanged page on the same domain tracked as a control.

Table 8: Brands in the study
BrandSectorMarket
Brand ACleaning business softwareNational (B2B)
Brand BCustom window treatmentsSingle US metro
Brand CShort-term rental managementSingle US metro
Brand DKitchen remodelingRegional (B2C)
Brand EWindow and door installationSingle US metro
Brand FDesign-build contractingSingle US metro
Brand GSecurity systems integrationSingle US metro

The change: A typical rewrite moved from a generic title to one built around the target phrase, for example “Top 10 [Category] Company Software” became “Best [Category] Business Software.” Several rewrites also changed other words (“Skilled” to “Reliable”, “Installers” to “Companies”), so the study cannot separate a verbatim prompt match from general rewording.

Prompts: Each test page was tracked against three prompts, and each control page against the same three:

  • Target: the prompt the new title was written to match. If titles matter, it should gain visibility.
  • Previous-title: the prompt the old title was built around. If titles matter, it should lose ground.
  • Control: a related prompt in neither title. It shows how a comparable query moves on its own.

Control prompts show how a related query moves, while control pages show whether a retrieval change is specific to the edited page.

Control pages: Control prompts show how a related query moves, not whether a retrieval change is specific to the edited page. For that, we tracked each brand's most-retrieved other page during the baseline (five homepages, one service page, and one blog post) against the same three prompts.

Platforms and timing. ChatGPT and Google AI Overviews were the focus, with Perplexity tracked alongside. Each prompt ran about 14 times per platform per window, so one response moves a visibility score by about 7 points

Table 9: Data collection windows
WindowDates
Baseline7 to 20 September 2026
Weeks 1 and 221 September to 4 October 2026
Weeks 3 and 45 to 18 October 2026

2. Metrics

All figures come from Peec AI, using the same prompts, platforms, and windows for test and control pages.

  • Visibility score: share of AI responses that mentioned the brand. For AI Overviews, only days an Overview appeared count (0 to 14 per prompt per window).
  • Combined visibility: mentions pooled across all runs on all three platforms, not an average of platform scores.
  • Retrievals: times the AI tool fetched the page, per prompt or across all tracked prompts. Peec AI counts a page as retrieved when it is fetched and enters the candidate set the model might use, a step before citation or mention.
  • Citation rate: average citations per retrieval, as reported by Peec AI.
  • Points: percentage-point change (48.9% to 28.6% is -20.3 points).
  • Multiplier (×): post-change retrievals relative to what the control trend predicted; ×1 means no change.
  • Odds ratio: change in the odds of a mention relative to the trend; below 1 means fewer mentions than expected.
  • p-value / 95% CI: the chance of an effect at least this large if titles had no effect (below 0.05 is significant), and the plausible range for the true multiplier.
  • Spearman ρ: rank correlation from -1 to 1; 0 means no relationship.
  • Measurable change: a target-prompt change at least 10 points larger than both controls, averaged over the two post-change windows.

3. Statistical approach

Both model types include a fixed effect for each prompt, a shared effect for each window, and terms for how the target and previous-title prompts moved after the change, with robust standard errors.

  • Retrievals: Poisson regressions, run separately per platform.
  • Visibility: binomial regressions on the number of runs that mentioned the brand.
  • Sensitivity: each retrieval model was re-run seven times, leaving out one brand each time.

Each platform model includes only prompts retrieved at least once on that platform (11 each for ChatGPT and AI Overviews), so intervals are approximate. Pooled models can detect far smaller effects than single prompts, which need a swing of about 43 points to reach p < 0.05. The ChatGPT visibility model found no detectable title effect (odds ratios 0.45 and 0.59 in the two post-change windows; p = 0.18 and 0.40). The same model could not be estimated reliably for AI Overviews.

4. Limitations

This sample shows no visibility effect from a title change within four weeks. It does not show that meta titles are irrelevant to AI search.

  • Small sample: seven pages and 21 prompts can detect large effects but not small ones.
  • Background movement: visibility fell across all prompt types during the study. Controls account for this, but noise of that size hides small effects.
  • Mixed industries: the sample is too small to compare national B2B software with single-metro home services.
  • Imperfect controls: most control pages are homepages, which AI tools retrieve for different reasons than service pages or list articles, and control-page volumes vary widely by brand.
  • Recrawl and display unconfirmed: we did not record when Google and Bing recrawled each page or whether Google displayed the new title. If recrawl lagged, weeks one and two understate the effect.
  • Short window: slower changes in how AI tools index and weigh pages could take longer than four weeks to appear.
  • Unreported metrics: sentiment and position were planned but are not in the dataset.

5. Data quality notes

These irregularities do not change the main findings but limit how far individual numbers can be trusted.

  • Brand F: all three prompts fell to near zero in weeks three and four while the page was still being retrieved, pointing to a tracking or brand-recognition issue rather than a real loss.
  • Perplexity zeros: in weeks three and four, 9 of 21 prompts showed exactly 0% visibility, against none at baseline, including some still being retrieved (Brand F’s previous-title prompt: 14 retrievals, 0% visibility). Control prompts were hit as often as target prompts, which points to a tracking or model change, not a title effect.
  • AI Overviews samples: individual scores rest on 1 to 14 appearances; Brand E’s target prompt (1 to 3 per window) is the most fragile. Two previous-title prompts that did not trigger Overviews at setup later did (57% and 60%; 50% then 0%), either because Google expanded Overviews or because of a tracking artifact.
  • Sparse control pages: Brand D’s control page was not retrieved for the study prompts in weeks three and four, and Brand E’s only once, so both add little to the comparison.
  • Run counts: for several prompt-windows, combined visibility cannot be rebuilt from three platform scores at 14 runs each, so some runs were likely missed. One retrieval figure differs by one between the export and the platform. Neither materially affects averages.

Sources

Are you winning AI visibility?

Or does it just look that way?

Most reports can’t tell the difference. Ours can.

ZeroClick Labs tests every change against control prompts and unchanged pages, measures each AI engine on its own, and puts retrievals and brand mentions side by side – so you know which wins are real before you throw another dollar at them.

Connect with us today, and let’s make changes that actually matter!

“Our agency had no idea how to approach AI visibility. ZeroClick only does this one thing so they actually know what works. Worth every penny just to not waste time figuring it out ourselves.” - Jay

Discover how ZeroClick Labs can strengthen your AI search presence.

Get More AI Insights