Job search
Frontend Performance Portfolio: Explain a Measurement, a Fix, and Its Limits
Explain a frontend performance change with traceable conditions, an honest result and a clear distinction between lab evidence and real-user outcomes.
A frontend performance portfolio should explain a measured problem, the change you made and what the evidence can support. Identify the page and test conditions, preserve the original observations and show the result even when it does not improve. Without the page, conditions and retained observations, a higher score leaves a reviewer unable to distinguish an engineering change from a favorable run.
This guide reviews current public Google and Chrome documentation checked October 22, 2026. Both completed cases use explicitly synthetic numbers and fictional pages. We execute only their arithmetic checks, not a Lighthouse audit, browser performance recording or user study. LandOffer publishes this guide and offers recent technical roles; use the role’s responsibilities to decide which engineering evidence belongs in your portfolio.

Choose a question before choosing a score
Start with a user-facing symptom and a specific technical boundary: the main image appears late on an article page, or a filter interaction delays visible feedback. Name the page or component you own and the evidence you are permitted to inspect. A local sample, a permitted client page and a production service with real traffic offer different evidence.
Google’s lab and field data explanation distinguishes controlled device and network conditions from data gathered during real visits. Lab work helps reproduce a problem and isolate a change. Field data reflects a distribution of devices, networks and behavior. Neither a local result nor a fictional dataset becomes real-user evidence by calling it representative.
Choose a metric that corresponds to the question. Largest Contentful Paint, or LCP, concerns when the largest eligible image or text block is rendered in the viewport. Interaction to Next Paint, or INP, concerns responsiveness across interactions during a visit. Cumulative Layout Shift concerns unexpected visual instability. A faster download, shorter JavaScript task and lower layout-shift value answer different questions.
The INP definition, updated September 2, 2025, includes input delay, processing and presentation delay. Timing a function alone does not establish INP. A navigation-only Lighthouse result also cannot tell you how every subsequent interaction behaves. Describe the interaction and observation method instead of relabeling a convenient number.
Preserve the conditions and the unedited observations
For an actual experiment, record the source revision, page route, content fixture, browser and tool versions, machine and operating system, viewport and device scale, network and CPU settings, cache state and interaction script. Include iteration order and retained reports or traces. If one detail was not captured, state the gap; do not fill it with a plausible machine description later.
Chrome’s Lighthouse scoring guide explains variability caused by underlying conditions, including device differences, extensions and changing content. Keep the individual observations beside the summary so a reviewer can see the spread, then state how you chose the reported value. A median can describe a small series, but it does not remove confounding factors or establish statistical significance.
Separate a lab scenario from a field population. If you have genuine field data, identify its source, reporting window, mobile or desktop segment and whether it represents a URL or an origin. Google’s guidance notes that the two scopes can differ. Missing URL-level data is an evidence limitation, not permission to assign an origin’s experience to your particular page.
A performance case also needs correctness checks. Does the page still show the intended content, handle errors and perform the same task? An optimization that removes the feature being measured is a different product change. State what you verified and what remains untested, including accessibility; this article is not an accessibility audit.
Worked case one: propose an earlier-discovered main image
Original synthetic case: a fictional public workshop listing renders its large opening image only after a client-side component initializes. The teaching hypothesis is that making the image discoverable in the initial HTML could reduce resource load delay. No actual listing, image request or browser trace was inspected for this case.
Google’s LCP optimization guide, updated March 31, 2025, recommends examining the entire loading process and making the LCP resource discoverable early. It separates time to first byte, resource load delay, resource load duration and element render delay. A smaller image alone may leave another delay unchanged.
Completed case record: the proposed intervention moves the same main-image reference into the initial document and avoids lazy-loading that above-the-fold LCP candidate. The image content, intended dimensions and page task remain the same. This is a proposed implementation boundary, not a claim that we changed a real repository or identified an actual page’s LCP element.
| Condition | Explicit hypothetical record |
|---|---|
| Page and content | Workshop listing revision A versus B, same invented title, image and listing entries |
| Viewport | Both series assume 390 × 844 CSS pixels and device scale 1 |
| Cache | Both assume a fresh browser cache for each observation; server-cache conditions held constant in the scenario |
| Network and CPU | Identical fixed profiles assumed across the two series; no actual profile, machine or tool version exists |
| Iterations | Five invented values per version, kept in their displayed order |
| Correctness boundary | Same image and listing task required; no actual rendering, functional or accessibility check performed |
| Synthetic observation | Before LCP, milliseconds | After LCP, milliseconds |
|---|---|---|
| 1 | 3200 | 2700 |
| 2 | 3400 | 2900 |
| 3 | 3300 | 2800 |
| 4 | 3500 | 3000 |
| 5 | 3100 | 2600 |
Finished calculation: sorting the before values yields a median of 3300 ms; the after median is 2800 ms. The absolute difference is 3300 − 2800 = 500 ms. Relative to the before median, 500 ÷ 3300 × 100 ≈ 15.15%. Those calculations are checked locally against the retained synthetic arrays. The percentage describes these invented medians, not measured page speed, a confidence interval or an improvement experienced by users.
Portfolio-ready fictional result paragraph: “In this synthetic workshop-listing exercise, I proposed exposing the same main image in the initial HTML to address delayed discovery. The teaching dataset’s five-value median changes from 3300 to 2800 ms. That supports the arithmetic illustration of a 500 ms difference; it does not validate the proposed browser mechanism. To claim an observed improvement, I would need traces identifying the LCP element and request timing, unchanged content checks and repeat observations under recorded conditions.”
Completed recommendation: retain early discovery as the next hypothesis to test on an owned or authorized fixture. Do not write “improved real-user LCP by 15%” or claim a shipped fix. The record explains the intervention clearly while preserving the difference between a proposed change, a synthetic calculation and an observed result.
Explain the mechanism without enlarging the claim
A useful portfolio connects the change to evidence from the relevant layer. If an actual waterfall shows late image discovery, preserve that request and its relationship to the document. If a trace shows a long task blocking interaction feedback, identify the affected work. Do not substitute a generic “optimized React” statement for a mechanism.
Explain alternative causes. A different image, warm cache, faster response or quieter machine could change a result without validating the intended intervention. Keep the page, dataset and route stable when possible; record deliberate differences. If image discovery, compression and rendering logic change together, a total difference cannot by itself tell the reviewer which change caused it.
Avoid equating the Lighthouse Performance score with a percentage of user experience. The scoring guide describes metric contributions to the aggregate score. Diagnostics and opportunities inform investigation, but their presence does not independently prove a particular time saving. Preserve the metric values and relevant trace rather than presenting only a colored badge.
For genuine field comparisons, deployment dates and reporting windows matter. A traffic mix can change, and a reporting period can include multiple versions. Explain those limitations alongside the result. Neither an improved lab series nor an origin-wide field value establishes conversion, revenue, retention or search-ranking impact.
Worked case two: a smaller image with no LCP improvement
Different original synthetic case: a fictional exhibition page compresses its main image while leaving a render dependency unresolved. Its resource downloads sooner, but the element still waits to render. This case demonstrates an unchanged outcome and a narrower recommendation; it is not an actual compression benchmark.
The LCP optimization guide explicitly describes how time saved in resource load duration can move into element render delay without improving total LCP. The invented representative breakdown below follows that reasoning. It does not copy measured values from Google’s example or assert that a trace was recorded here.
| Synthetic LCP subpart | Before, milliseconds | After, milliseconds |
|---|---|---|
| Time to first byte | 600 | 600 |
| Resource load delay | 800 | 800 |
| Resource load duration | 1000 | 700 |
| Element render delay | 1000 | 1300 |
| Total | 3400 | 3400 |
Finished breakdown: before is 600 + 800 + 1000 + 1000 = 3400 ms. After is 600 + 800 + 700 + 1300 = 3400 ms. Download duration falls by 300 ms while render delay grows by 300 ms. Total change is zero. This is an additive illustration for an image-based LCP resource, not a general formula for unrelated timing measurements.
The second retained dataset keeps five invented total-LCP values for each version: before 3400, 3500, 3300, 3400, 3600; after 3400, 3300, 3500, 3400, 3600. Both medians are 3400 ms, with ranges from 3300 to 3600 ms. The representative breakdown and the five-value series are separate teaching records; we do not pretend to possess subpart measurements for every observation.

Portfolio-ready fictional finding: “The synthetic compression case shortens resource load duration but leaves total LCP unchanged at 3400 ms in the representative breakdown. The five-value median is also unchanged. I would retain the smaller resource as a candidate change only after checking visual quality and other costs; the next LCP investigation should examine why rendering waits after the image is ready.”
Completed decision record: do not headline this case as a successful LCP optimization. Preserve the no-improvement result, distinguish potential transfer-size benefit from LCP and investigate the render dependency before proposing another fix. No byte-size reduction or visual-quality result was measured here, so neither belongs in an achievement bullet.
This case offers useful engineering reasoning because it narrows the next question. Preserving the unchanged total shows why the next investigation concerns rendering rather than another claim about download duration. In a real project, retain the actual trace, change and correctness results that support that decision.
Package a case that a reviewer can examine
Put the question, scope, baseline, intervention, observations, calculation and limitation in one readable case study. Link permitted code changes and raw reports where they genuinely exist. Give a reviewer enough context to reproduce the method without requiring private credentials, customer data or a production deployment.
Label synthetic exercises as synthetic in the title or opening of the case, not only in a distant footnote. If you later perform a real experiment, retain its new evidence separately and update the claim accordingly. Do not rename the arithmetic output “Lighthouse results”; the checker establishes calculations, not performance.
End the case with the next technical question and the observation that makes it worth investigating. Keep an unchanged result when it improves the diagnosis. As you review recent LandOffer roles, select a performance case that shows the system you understand, the work you actually performed and the limits you can explain.
Sources and Further Reading
Public primary documentation checked October 22, 2026. Examples and numeric arrays are original synthetic teaching material; only arithmetic was executed.


